Scikit-Learn 库中的分类模型及其导出到 ONNX
MetaTrader 5
—
示例
| 6 六月 2024, 11:10
278
0
技术的发展带来了一种构建数据处理算法的全新方法。以前，为了解决每一个特定任务，都需要明确的形式化和相应算法的开发。
在机器学习中，计算机会学习自行找到处理数据的最佳方法。机器学习模型可以成功地解决分类任务（其中有一组固定的类，目标是找到属于每个类的给定特征集的概率）和回归任务（其中目标是基于给定特征集估计目标变量的数值）。可以基于这些基本组件构建更复杂的数据处理模型。
Scikit-learn 库为分类和回归提供了多种工具。具体方法和模型的选择取决于数据的特点，因为不同的方法可能具有不同的有效性，并根据任务提供不同的结果。
在新闻稿 “ONNX Runtime 现已开源”中，声明了 ONNX Runtime 还支持 ONNX-ML 配置文件：
ONNX Runtime是第一个完全支持 ONNX 1.2及更高版本（包括 ONNX-ML 配置文件）的公开推理引擎。
ONNX-ML 配置文件是 ONNX 的一部分，专为机器学习 (ML) 模型设计。它旨在以方便的格式描述和表示各种类型的 ML 模型，例如分类、回归、聚类等，可以在支持 ONNX 的各种平台和环境中使用。ONNX-ML 配置文件简化了机器学习模型的传输、部署和执行，使其更易于访问和移植。
在本文中，我们将探讨 Scikit-learn 包中的所有分类模型在解决 Fisher 鸢尾花分类任务中的应用。我们还将尝试将这些模型转换为 ONNX 格式，并在 MQL5 程序中使用生成的模型。
此外，我们将在完整的鸢尾花数据集上比较原始模型与其 ONNX 版本的准确性。
目录
- 1.费舍尔的鸢尾花
- 2.分类模型
Scikit-learn 分类器列表
模型的不同输出表示 iris.mqh - 2.1.SVC Classifier
2.1.1.创建 SVC Classifier 模型的代码
2.1.2.用于处理 SVC Classifier 模型的 MQL5 代码
2.1.3.SVC Classifier 模型的 ONNX 表示 - 2.2.LinearSVC Classifier
2.2.1.创建 LinearSVC Classifier 模型的代码
2.2.2.用于处理 LinearSVC Classifier 模型的 MQL5 代码
2.2.3.LinearSVC Classifier 模型的 ONNX 表示 - 2.3.NuSVC Classifier
2.3.1.创建 NuSVC Classifier 模型的代码
2.3.2.用于处理 NuSVC Classifier 模型的 MQL5 代码
2.3.3.NuSVC Classifier 模型的 ONNX 表示 - 2.4.Radius Neighbors Classifier
2.4.1.创建 Radius Neighbors Classifier 模型的代码
2.4.2.用于处理 Radius Neighbors Classifier 模型的 MQL5 代码
2.3.3.Radius Neighbors Classifier 模型的 ONNX 表示 - 2.5.Ridge Classifier
2.5.1.创建 Ridge Classifier 模型的代码
2.5.2.用于处理 Ridge Classifier 模型的 MQL5 代码
2.5.3.Ridge Classifier 模型的 ONNX 表示 - 2.6.RidgeClassifierCV
2.6.1.创建 Ridge ClassifierCV 模型的代码
2.6.2.用于处理 Ridge ClassifierCV 模型的 MQL5 代码
2.6.3.Ridge ClassifierCV 模型的 ONNX 表示 - 2.7.Random Forest Classifier
2.7.1.创建 Random Forest Classifier 模型的代码
2.7.2.用于处理 Random Forest Classifier 模型的 MQL5 代码
2.7.3.Random Forest Classifier 模型的 ONNX 表示 - 2.8.Gradient Boosting Classifier
2.8.1.创建 Gradient Boosting Classifier 模型的代码
2.8.2.用于处理 Gradient Boosting Classifier 模型的 MQL5 代码
2.8.3.Gradient Boosting Classifier 模型的 ONNX 表示 - 2.9.Adaptive Boosting Classifier
2.9.1.创建 Adaptive Boosting Classifier 模型的代码
2.9.2.用于处理 Adaptive Boosting Classifier 模型的 MQL5 代码
2.9.3.Adaptive Boosting Classifier 模型的 ONNX 表示 - 2.10.Bootstrap Aggregating Classifier
2.10.1.创建 Bootstrap Aggregating Classifier 模型的代码
2.10.2.用于处理 Bootstrap Aggregating Classifier 模型的 MQL5 代码
2.10.3.Bootstrap Aggregating Classifier 模型的 ONNX 表示 - 2.11.K-Nearest Neighbors (K-NN) Classifier
2.11.1.创建 K-Nearest Neighbors (K-NN) Classifier 模型的代码
2.11.2.用于处理 K-Nearest Neighbors (K-NN) Classifier 模型的 MQL5 代码
2.11.3.K-Nearest Neighbors (K-NN) Classifier 模型的 ONNX 表示 - 2.12.Decision Tree Classifier
2.12.1.创建 Decision Tree Classifier 模型的代码
2.12.2.用于处理 Decision Tree Classifier 模型的 MQL5 代码
2.12.3.Decision Tree Classifier 模型的 ONNX 表示 - 2.13.Logistic Regression Classifier
2.13.1.创建 Logistic Regression Classifier 模型的代码
2.13.2.用于处理 Logistic Regression Classifier 模型的 MQL5 代码
2.13.3.Logistic Regression Classifier 模型的 ONNX 表示 - 2.14.LogisticRegressionCV Classifier
2.14.1.创建 LogisticRegressionCV Classifier 模型的代码
2.14.2.用于处理 LogisticRegressionCV Classifier 模型的 MQL5 代码
2.14.3.LogisticRegressionCV Classifier 模型的 ONNX 表示 - 2.15.Passive-Aggressive (PA) Classifier
2.15.1.创建 Passive-Aggressive (PA) Classifier 模型的代码
2.15.2.用于处理 Passive-Aggressive (PA) Classifier 模型的 MQL5 代码
2.15.3.Passive-Aggressive (PA) Classifier 模型的 ONNX 表示 - 2.16.Perceptron Classifier
2.16.1.创建 Perceptron Classifier 模型的代码
2.16.2.用于处理 Perceptron Classifier 模型的 MQL5 代码
2.16.3.Perceptron Classifier 模型的 ONNX 表示 - 2.17.Stochastic Gradient Descent Classifier
2.17.1.创建 Stochastic Gradient Descent Classifier 模型的代码
2.17.2.用于处理 Stochastic Gradient Descent Classifier 模型的 MQL5 代码
2.17.3.Stochastic Gradient Descent Classifier 模型的 ONNX 表示 - 2.18.Gaussian Naive Bayes (GNB) Classifier
2.18.1.创建 Gaussian Naive Bayes (GNB) Classifier 模型的代码
2.18.2.用于处理 Gaussian Naive Bayes (GNB) Classifier 模型的 MQL5 代码
2.18.3.Gaussian Naive Bayes (GNB) Classifier 模型的 ONNX 表示 - 2.19.Multinomial Naive Bayes (MNB) Classifier
2.19.1.创建 Multinomial Naive Bayes (MNB) Classifier 模型的代码
2.19.2.用于处理 Multinomial Naive Bayes (MNB) Classifier 模型的 MQL5 代码
2.19.3.Multinomial Naive Bayes (MNB) Classifier 模型的 ONNX 表示 - 2.20.Complement Naive Bayes (CNB) Classifier
2.20.1.创建 Complement Naive Bayes (CNB) Classifier 模型的代码
2.20.2.用于处理 Complement Naive Bayes (CNB) Classifier 模型的 MQL5 代码
2.20.3.Complement Naive Bayes (CNB) Classifier 模型的 ONNX 表示 - 2.21.Bernoulli Naive Bayes (BNB) Classifier
2.21.1.创建 Bernoulli Naive Bayes (BNB) Classifier 模型的代码
2.21.2.用于处理 Bernoulli Naive Bayes (BNB) Classifier 的 MQL5 代码
2.21.3.Bernoulli Naive Bayes (BNB) Classifier 模型的 ONNX 表示 - 2.22.Multilayer Perceptron Classifier
2.22.1.创建 Multilayer Perceptron Classifier 模型的代码
2.22.2.用于处理 Multilayer Perceptron Classifier 模型的 MQL5 代码
2.22.3.Multilayer Perceptron Classifier 模型的 ONNX 表示 - 2.23.Linear Discriminant Analysis (LDA) Classifier
2.23.1.创建 Linear Discriminant Analysis (LDA) Classifier 模型的代码
2.23.2.用于处理 Linear Discriminant Analysis (LDA) Classifier 模型的 MQL5 代码
2.23.3.Linear Discriminant Analysis (LDA) Classifier 模型的 ONNX 表示 - 2.24.Hist Gradient Boosting
2.24.1.创建 Histogram-Based Gradient Boosting Classifier 模型的代码
2.24.2.用于处理 Histogram-Based Gradient Boosting Classifier 模型的 MQL5 代码
2.24.3.Histogram-Based Gradient Boosting Classifier 模型的 ONNX 表示 - 2.25。CategoricalNB Classifier
2.25.1.创建 CategoricalNB Classifier 模型的代码
2.25.2.用于处理 CategoricalNB Classifier 模型的 MQL5 代码
2.25.3.CategoricalNB Classifier 模型的 ONNX 表示 - 2.26.ExtraTreeClassifier
2.26.1.创建 ExtraTreeClassifier 模型的代码
2.26.2.用于处理 ExtraTreeClassifier 模型的 MQL5 代码
2.26.3.ExtraTreeClassifier 模型的 ONNX 表示 - 2.27.ExtraTreesClassifier
2.27.1.创建 ExtraTreesClassifier 模型的代码
2.27.2.用于处理 ExtraTreesClassifier 模型的 MQL5 代码
2.27.3.ExtraTreesClassifier 模型的 ONNX 表示 - 2.28.比较所有模型的准确率
2.28.1.计算所有模型并建立准确率比较图表的代码
2.28.2.用于执行所有 ONNX 模型的 MQL5 代码 - 2.29.无法转换为 ONNX 的 Scikit-Learn 分类模型
- 2.29.1.DummyClassifier
2.29.1.1.创建 DummyClassifier 模型的代码 - 2.29.2.GaussianProcessClassifier
2.29.2.1.创建 GaussianProcessClassifier 模型的代码 - 2.29.3.LabelPropagation Classifier
2.29.3.1.创建 LabelPropagationClassifier 模型的代码 - 2.29.4.LabelSpreading Classifier
2.29.4.1.创建 LabelSpreadingClassifier 模型的代码 - 2.29.5.NearestCentroid Classifier
2.29.5.1.创建 NearestCentroid 模型的代码 - 2.29.6.Quadratic Discriminant Analysis Classifier
2.29.6.1.创建 Quadratic Discriminant Analysis 模型的代码 - 结论
1.费舍尔的（Fisher's）鸢尾花
鸢尾花数据集是机器学习领域最著名和应用最广泛的数据集之一。它于 1936 年由统计学家和生物学家 R.A.Fisher 首次提出，从此成为分类任务的经典数据集。
鸢尾花数据集包括三种鸢尾花（山鸢尾、维吉尼亚鸢尾和杂色鸢尾）的萼片和花瓣的测量数据。
图1.山鸢尾
图 2.维吉尼亚鸢尾
图 3.杂色鸢尾
鸢尾花数据集包含 150 个鸢尾花实例，三个品种各有 50 个实例。每个实例有四个数值特征（以厘米为单位）：
- 花萼长度
- 花萼宽度
- 花瓣长度
- 花瓣宽度
每个实例还具有相应的类，表示鸢尾花种类（山鸢尾、维吉尼亚鸢尾或杂色鸢尾）。这种分类属性使鸢尾花数据集成为分类和聚类等机器学习任务的理想数据集。
MetaEditor 允许使用 Python 脚本。要创建 Python 脚本，请从 MetaEditor 中的“文件”菜单中选择“新建”，然后会出现一个用于选择要创建对象的对话框（见图 4）。
图 4.在 MQL5 向导中创建 Python 脚本 - 步骤 1
接下来，为脚本提供一个名称，例如“IRIS.py”（见图 5）。
图 5.在 MQL5 向导中创建 Python 脚本 - 步骤 2 - 脚本名称
之后，您可以指定将使用哪些库。在我们的例子中，我们将这些字段留空（见图 6）。
图6：在 MQL5 向导中创建 Python 脚本 - 步骤 3
开始分析鸢尾花数据集的一种方法是将数据可视化，图形表示可以让我们更好地理解数据的结构和特征之间的关系。
例如，您可以创建散点图来查看不同种类的鸢尾花在特征空间中的分布情况。
Python脚本代码：
# The script shows the scatter plot of the Iris dataset features # Copyright 2023, MetaQuotes Ltd. # https://mql5.com import matplotlib.pyplot as plt from sklearn import datasets # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # extract sepal length and sepal width (the first two features) sepal_length = X[:, 0] sepal_width = X[:, 1] # create a scatter plot plt.figure(figsize=(8, 6)) plt.scatter(sepal_length, sepal_width, c=y, cmap=plt.cm.Set1, edgecolor='k') plt.xlabel('Sepal Length (cm)') plt.ylabel('Sepal Width (cm)') plt.title('Scatter Plot for Sepal Length and Sepal Width') plt.colorbar(label='Iris Species', ticks=[0, 1, 2]) plt.show() # save the scatter plot to a file (optional) # plt.savefig('scatter_plot_sepal_length_width.png') # Extract petal length and petal width (the third and fourth features) petal_length = X[:, 2] petal_width = X[:, 3] # create a scatter plot plt.figure(figsize=(8, 6)) plt.scatter(petal_length, petal_width, c=y, cmap=plt.cm.Set1, edgecolor='k') plt.xlabel('Petal Length (cm)') plt.ylabel('Petal Width (cm)') plt.title('Scatter Plot for Petal Length and Petal Width') plt.colorbar(label='Iris Species', ticks=[0, 1, 2]) plt.show() # save the scatter plot to a file (optional) # plt.savefig('scatter_plot_petal_length_width.png')
要运行此脚本，您需要将其复制到 MetaEditor 中（见图 7）并单击“编译”。
图 7：MetaEditor 中的 IRIS.py 脚本
此后，图表将出现在屏幕上：
图 8：MetaEditor 中的 IRIS.py 脚本，带有花萼长度/花萼宽度图
图 9：MetaEditor 中的 IRIS.py 脚本，带有花瓣长度/花瓣宽度图
让我们仔细看看它们。
图 10：花萼长度与花萼宽度的散点图
在该图中，我们可以看到不同鸢尾花品种根据花萼长度和花萼宽度是如何分布的。我们可以观察到，与其他两个物种相比，山鸢尾的花萼通常更短更宽。
图 11：花瓣长度与花瓣宽度的散点图
在该图中，我们可以看到不同种类的鸢尾花是如何根据花瓣长度和花瓣宽度进行分布的。我们可以注意到，山鸢尾的花瓣最短且最窄，维吉尼亚鸢尾的花瓣最长且最宽，而杂色鸢尾则介于两者之间。
鸢尾花数据集是训练和测试机器学习模型的理想数据集，我们将用它来分析机器学习模型对于分类任务的有效性。
2.分类模型
分类是机器学习的基本任务之一，其目标是根据某些特征将数据分为不同的类别或种类。
让我们探索 scikit-learn 包中的主要机器学习模型。
Scikit-learn 分类器列表
要显示 scikit-learn 中可用的分类器列表，可以使用以下脚本：
# ScikitLearnClassifiers.py # The script lists all the classification algorithms available in scikit-learn # Copyright 2023, MetaQuotes Ltd. # https://mql5.com # print Python version from platform import python_version print("The Python version is ", python_version()) # print scikit-learn version import sklearn print('The scikit-learn version is {}.'.format(sklearn.__version__)) # print scikit-learn classifiers from sklearn.utils import all_estimators classifiers = all_estimators(type_filter='classifier') for index, (name, ClassifierClass) in enumerate(classifiers, start=1): print(f"Classifier {index}: {name}")
输出：
Python The Python version is 3.10.0
Python The scikit-learn version is 1.2.2.
Python Classifier 1:AdaBoostClassifier
Python Classifier 2:BaggingClassifier
Python Classifier 3:BernoulliNB
Python Classifier 4:CalibratedClassifierCV
Python Classifier 5:CategoricalNB
Python Classifier 6:ClassifierChain
Python Classifier 7:ComplementNB
Python Classifier 8:DecisionTreeClassifier
Python Classifier 9:DummyClassifier
Python Classifier 10:ExtraTreeClassifier
Python Classifier 11:ExtraTreesClassifier
Python Classifier 12:GaussianNB
Python Classifier 13:GaussianProcessClassifier
Python Classifier 14:GradientBoostingClassifier
Python Classifier 15:HistGradientBoostingClassifier
Python Classifier 16:KNeighborsClassifier
Python Classifier 17:LabelPropagation
Python Classifier 18:LabelSpreading
Python Classifier 19:LinearDiscriminantAnalysis
Python Classifier 20:LinearSVC
Python Classifier 21:LogisticRegression
Python Classifier 22:LogisticRegressionCV
Python Classifier 23:MLPClassifier
Python Classifier 24:MultiOutputClassifier
Python Classifier 25:MultinomialNB
Python Classifier 26:NearestCentroid
Python Classifier 27:NuSVC
Python Classifier 28:OneVsOneClassifier
Python Classifier 29:OneVsRestClassifier
Python Classifier 30:OutputCodeClassifier
Python Classifier 31:PassiveAggressiveClassifier
Python Classifier 32:Perceptron
Python Classifier 33:QuadraticDiscriminantAnalysis
Python Classifier 34:RadiusNeighborsClassifier
Python Classifier 35:RandomForestClassifier
Python Classifier 36:RidgeClassifier
Python Classifier 37:RidgeClassifierCV
Python Classifier 38:SGDClassifier
Python Classifier 39:SVC
Python Classifier 40:StackingClassifier
Python Classifier 41:VotingClassifier
Python The scikit-learn version is 1.2.2.
Python Classifier 1:AdaBoostClassifier
Python Classifier 2:BaggingClassifier
Python Classifier 3:BernoulliNB
Python Classifier 4:CalibratedClassifierCV
Python Classifier 5:CategoricalNB
Python Classifier 6:ClassifierChain
Python Classifier 7:ComplementNB
Python Classifier 8:DecisionTreeClassifier
Python Classifier 9:DummyClassifier
Python Classifier 10:ExtraTreeClassifier
Python Classifier 11:ExtraTreesClassifier
Python Classifier 12:GaussianNB
Python Classifier 13:GaussianProcessClassifier
Python Classifier 14:GradientBoostingClassifier
Python Classifier 15:HistGradientBoostingClassifier
Python Classifier 16:KNeighborsClassifier
Python Classifier 17:LabelPropagation
Python Classifier 18:LabelSpreading
Python Classifier 19:LinearDiscriminantAnalysis
Python Classifier 20:LinearSVC
Python Classifier 21:LogisticRegression
Python Classifier 22:LogisticRegressionCV
Python Classifier 23:MLPClassifier
Python Classifier 24:MultiOutputClassifier
Python Classifier 25:MultinomialNB
Python Classifier 26:NearestCentroid
Python Classifier 27:NuSVC
Python Classifier 28:OneVsOneClassifier
Python Classifier 29:OneVsRestClassifier
Python Classifier 30:OutputCodeClassifier
Python Classifier 31:PassiveAggressiveClassifier
Python Classifier 32:Perceptron
Python Classifier 33:QuadraticDiscriminantAnalysis
Python Classifier 34:RadiusNeighborsClassifier
Python Classifier 35:RandomForestClassifier
Python Classifier 36:RidgeClassifier
Python Classifier 37:RidgeClassifierCV
Python Classifier 38:SGDClassifier
Python Classifier 39:SVC
Python Classifier 40:StackingClassifier
Python Classifier 41:VotingClassifier
为了方便起见，此分类器列表以不同的颜色突出显示。需要基础分类器的模型以黄色突出显示，而其他模型可以独立使用。
展望未来，值得注意的是，绿色模型已成功导出为 ONNX 格式，而红色模型在当前版本的 scikit-learn 1.2.2 中转换时遇到错误。
模型中输出数据的不同表示
需要注意的是，不同的模型对输出数据的表示不同，因此在使用转换为 ONNX 的模型时应该注意。
对于 Fisher 的鸢尾花分类任务，所有这些模型的输入张量都具有相同的格式：
Information about input tensors in ONNX:
1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
ONNX 模型的输出张量不同。
1.不需要后处理的模型：
- SVC Classifier;
- LinearSVC Classifier;
- NuSVC Classifier;
- Radius Neighbors Classifier;
- Ridge Classifier;
- Ridge Classifier CV.
Information about output tensors in ONNX:
1.Name: label, Data Type: tensor(int64), Shape: [None]
2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
1.Name: label, Data Type: tensor(int64), Shape: [None]
2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
这些模型在第一个输出整数张量“label”中明确返回结果（类编号），而无需后期处理。
2.结果需要后处理的模型：
- Random Forest Classifier;
- Gradient Boosting Classifier;
- AdaBoost Classifier;
- Bagging Classifier;
- K-NN_Classifier;
- Decision Tree Classifier;
- Logistic Regression Classifier;
- Logistic Regression CV Classifier;
- Passive-Aggressive Classifier;
- Perceptron Classifier;
- SGD Classifier;
- Gaussian Naive Bayes Classifier;
- Multinomial Naive Bayes Classifier;
- Complement Naive Bayes Classifier;
- Bernoulli Naive Bayes Classifier;
- Multilayer Perceptron Classifier;
- Linear Discriminant Analysis Classifier;
- Hist Gradient Boosting Classifier;
- Categorical Naive Bayes Classifier;
- ExtraTree Classifier;
- ExtraTrees Classifier.
Information about output tensors in ONNX:
1.Name: output_label, Data Type: tensor(int64), Shape: [None]
2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
1.Name: output_label, Data Type: tensor(int64), Shape: [None]
2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
这些模型返回类别列表以及属于每个类别的概率。
在这些情况下，为了获得结果，需要进行后处理，例如 seq(map(int64, tensor(float))（找到概率最高的元素）。
因此，在使用 ONNX 模型时必须注意并考虑这些方面。2.28.2中的脚本给出了不同结果处理的示例。
iris.mqh
为了在 MQL5 中对完整的鸢尾花数据集进行模型测试，需要进行数据准备。为此，将使用函数 PrepareIrisDataset()。
将这些函数移至 iris.mqh 文件会比较方便。
//+------------------------------------------------------------------+ //| Iris.mqh | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" //+------------------------------------------------------------------+ //| Structure for the IRIS Dataset sample | //+------------------------------------------------------------------+ struct sIRISsample { int sample_id; // sample id (1-150) double features[4]; // SepalLengthCm,SepalWidthCm,PetalLengthCm,PetalWidthCm string class_name; // class ("Iris-setosa","Iris-versicolor","Iris-virginica") int class_id; // class id (0,1,2), calculated by function IRISClassID }; //--- Iris dataset sIRISsample ExtIRISDataset[]; int Exttotal=0; //+------------------------------------------------------------------+ //| Returns class id by class name | //+------------------------------------------------------------------+ int IRISClassID(string class_name) { //--- if(class_name=="Iris-setosa") return(0); else if(class_name=="Iris-versicolor") return(1); else if(class_name=="Iris-virginica") return(2); //--- return(-1); } //+------------------------------------------------------------------+ //| AddSample | //+------------------------------------------------------------------+ bool AddSample(const int Id,const double SepalLengthCm,const double SepalWidthCm,const double PetalLengthCm,const double PetalWidthCm, const string Species) { //--- ExtIRISDataset[Exttotal].sample_id=Id; //--- ExtIRISDataset[Exttotal].features[0]=SepalLengthCm; ExtIRISDataset[Exttotal].features[1]=SepalWidthCm; ExtIRISDataset[Exttotal].features[2]=PetalLengthCm; ExtIRISDataset[Exttotal].features[3]=PetalWidthCm; //--- ExtIRISDataset[Exttotal].class_name=Species; ExtIRISDataset[Exttotal].class_id=IRISClassID(Species); //--- Exttotal++; //--- return(true); } //+------------------------------------------------------------------+ //| Prepare Iris Dataset | //+------------------------------------------------------------------+ bool PrepareIrisDataset(sIRISsample &iris_samples[]) { ArrayResize(ExtIRISDataset,150); Exttotal=0; //--- AddSample(1,5.1,3.5,1.4,0.2,"Iris-setosa"); AddSample(2,4.9,3.0,1.4,0.2,"Iris-setosa"); AddSample(3,4.7,3.2,1.3,0.2,"Iris-setosa"); AddSample(4,4.6,3.1,1.5,0.2,"Iris-setosa"); AddSample(5,5.0,3.6,1.4,0.2,"Iris-setosa"); AddSample(6,5.4,3.9,1.7,0.4,"Iris-setosa"); AddSample(7,4.6,3.4,1.4,0.3,"Iris-setosa"); AddSample(8,5.0,3.4,1.5,0.2,"Iris-setosa"); AddSample(9,4.4,2.9,1.4,0.2,"Iris-setosa"); AddSample(10,4.9,3.1,1.5,0.1,"Iris-setosa"); AddSample(11,5.4,3.7,1.5,0.2,"Iris-setosa"); AddSample(12,4.8,3.4,1.6,0.2,"Iris-setosa"); AddSample(13,4.8,3.0,1.4,0.1,"Iris-setosa"); AddSample(14,4.3,3.0,1.1,0.1,"Iris-setosa"); AddSample(15,5.8,4.0,1.2,0.2,"Iris-setosa"); AddSample(16,5.7,4.4,1.5,0.4,"Iris-setosa"); AddSample(17,5.4,3.9,1.3,0.4,"Iris-setosa"); AddSample(18,5.1,3.5,1.4,0.3,"Iris-setosa"); AddSample(19,5.7,3.8,1.7,0.3,"Iris-setosa"); AddSample(20,5.1,3.8,1.5,0.3,"Iris-setosa"); AddSample(21,5.4,3.4,1.7,0.2,"Iris-setosa"); AddSample(22,5.1,3.7,1.5,0.4,"Iris-setosa"); AddSample(23,4.6,3.6,1.0,0.2,"Iris-setosa"); AddSample(24,5.1,3.3,1.7,0.5,"Iris-setosa"); AddSample(25,4.8,3.4,1.9,0.2,"Iris-setosa"); AddSample(26,5.0,3.0,1.6,0.2,"Iris-setosa"); AddSample(27,5.0,3.4,1.6,0.4,"Iris-setosa"); AddSample(28,5.2,3.5,1.5,0.2,"Iris-setosa"); AddSample(29,5.2,3.4,1.4,0.2,"Iris-setosa"); AddSample(30,4.7,3.2,1.6,0.2,"Iris-setosa"); AddSample(31,4.8,3.1,1.6,0.2,"Iris-setosa"); AddSample(32,5.4,3.4,1.5,0.4,"Iris-setosa"); AddSample(33,5.2,4.1,1.5,0.1,"Iris-setosa"); AddSample(34,5.5,4.2,1.4,0.2,"Iris-setosa"); AddSample(35,4.9,3.1,1.5,0.2,"Iris-setosa"); AddSample(36,5.0,3.2,1.2,0.2,"Iris-setosa"); AddSample(37,5.5,3.5,1.3,0.2,"Iris-setosa"); AddSample(38,4.9,3.6,1.4,0.1,"Iris-setosa"); AddSample(39,4.4,3.0,1.3,0.2,"Iris-setosa"); AddSample(40,5.1,3.4,1.5,0.2,"Iris-setosa"); AddSample(41,5.0,3.5,1.3,0.3,"Iris-setosa"); AddSample(42,4.5,2.3,1.3,0.3,"Iris-setosa"); AddSample(43,4.4,3.2,1.3,0.2,"Iris-setosa"); AddSample(44,5.0,3.5,1.6,0.6,"Iris-setosa"); AddSample(45,5.1,3.8,1.9,0.4,"Iris-setosa"); AddSample(46,4.8,3.0,1.4,0.3,"Iris-setosa"); AddSample(47,5.1,3.8,1.6,0.2,"Iris-setosa"); AddSample(48,4.6,3.2,1.4,0.2,"Iris-setosa"); AddSample(49,5.3,3.7,1.5,0.2,"Iris-setosa"); AddSample(50,5.0,3.3,1.4,0.2,"Iris-setosa"); AddSample(51,7.0,3.2,4.7,1.4,"Iris-versicolor"); AddSample(52,6.4,3.2,4.5,1.5,"Iris-versicolor"); AddSample(53,6.9,3.1,4.9,1.5,"Iris-versicolor"); AddSample(54,5.5,2.3,4.0,1.3,"Iris-versicolor"); AddSample(55,6.5,2.8,4.6,1.5,"Iris-versicolor"); AddSample(56,5.7,2.8,4.5,1.3,"Iris-versicolor"); AddSample(57,6.3,3.3,4.7,1.6,"Iris-versicolor"); AddSample(58,4.9,2.4,3.3,1.0,"Iris-versicolor"); AddSample(59,6.6,2.9,4.6,1.3,"Iris-versicolor"); AddSample(60,5.2,2.7,3.9,1.4,"Iris-versicolor"); AddSample(61,5.0,2.0,3.5,1.0,"Iris-versicolor"); AddSample(62,5.9,3.0,4.2,1.5,"Iris-versicolor"); AddSample(63,6.0,2.2,4.0,1.0,"Iris-versicolor"); AddSample(64,6.1,2.9,4.7,1.4,"Iris-versicolor"); AddSample(65,5.6,2.9,3.6,1.3,"Iris-versicolor"); AddSample(66,6.7,3.1,4.4,1.4,"Iris-versicolor"); AddSample(67,5.6,3.0,4.5,1.5,"Iris-versicolor"); AddSample(68,5.8,2.7,4.1,1.0,"Iris-versicolor"); AddSample(69,6.2,2.2,4.5,1.5,"Iris-versicolor"); AddSample(70,5.6,2.5,3.9,1.1,"Iris-versicolor"); AddSample(71,5.9,3.2,4.8,1.8,"Iris-versicolor"); AddSample(72,6.1,2.8,4.0,1.3,"Iris-versicolor"); AddSample(73,6.3,2.5,4.9,1.5,"Iris-versicolor"); AddSample(74,6.1,2.8,4.7,1.2,"Iris-versicolor"); AddSample(75,6.4,2.9,4.3,1.3,"Iris-versicolor"); AddSample(76,6.6,3.0,4.4,1.4,"Iris-versicolor"); AddSample(77,6.8,2.8,4.8,1.4,"Iris-versicolor"); AddSample(78,6.7,3.0,5.0,1.7,"Iris-versicolor"); AddSample(79,6.0,2.9,4.5,1.5,"Iris-versicolor"); AddSample(80,5.7,2.6,3.5,1.0,"Iris-versicolor"); AddSample(81,5.5,2.4,3.8,1.1,"Iris-versicolor"); AddSample(82,5.5,2.4,3.7,1.0,"Iris-versicolor"); AddSample(83,5.8,2.7,3.9,1.2,"Iris-versicolor"); AddSample(84,6.0,2.7,5.1,1.6,"Iris-versicolor"); AddSample(85,5.4,3.0,4.5,1.5,"Iris-versicolor"); AddSample(86,6.0,3.4,4.5,1.6,"Iris-versicolor"); AddSample(87,6.7,3.1,4.7,1.5,"Iris-versicolor"); AddSample(88,6.3,2.3,4.4,1.3,"Iris-versicolor"); AddSample(89,5.6,3.0,4.1,1.3,"Iris-versicolor"); AddSample(90,5.5,2.5,4.0,1.3,"Iris-versicolor"); AddSample(91,5.5,2.6,4.4,1.2,"Iris-versicolor"); AddSample(92,6.1,3.0,4.6,1.4,"Iris-versicolor"); AddSample(93,5.8,2.6,4.0,1.2,"Iris-versicolor"); AddSample(94,5.0,2.3,3.3,1.0,"Iris-versicolor"); AddSample(95,5.6,2.7,4.2,1.3,"Iris-versicolor"); AddSample(96,5.7,3.0,4.2,1.2,"Iris-versicolor"); AddSample(97,5.7,2.9,4.2,1.3,"Iris-versicolor"); AddSample(98,6.2,2.9,4.3,1.3,"Iris-versicolor"); AddSample(99,5.1,2.5,3.0,1.1,"Iris-versicolor"); AddSample(100,5.7,2.8,4.1,1.3,"Iris-versicolor"); AddSample(101,6.3,3.3,6.0,2.5,"Iris-virginica"); AddSample(102,5.8,2.7,5.1,1.9,"Iris-virginica"); AddSample(103,7.1,3.0,5.9,2.1,"Iris-virginica"); AddSample(104,6.3,2.9,5.6,1.8,"Iris-virginica"); AddSample(105,6.5,3.0,5.8,2.2,"Iris-virginica"); AddSample(106,7.6,3.0,6.6,2.1,"Iris-virginica"); AddSample(107,4.9,2.5,4.5,1.7,"Iris-virginica"); AddSample(108,7.3,2.9,6.3,1.8,"Iris-virginica"); AddSample(109,6.7,2.5,5.8,1.8,"Iris-virginica"); AddSample(110,7.2,3.6,6.1,2.5,"Iris-virginica"); AddSample(111,6.5,3.2,5.1,2.0,"Iris-virginica"); AddSample(112,6.4,2.7,5.3,1.9,"Iris-virginica"); AddSample(113,6.8,3.0,5.5,2.1,"Iris-virginica"); AddSample(114,5.7,2.5,5.0,2.0,"Iris-virginica"); AddSample(115,5.8,2.8,5.1,2.4,"Iris-virginica"); AddSample(116,6.4,3.2,5.3,2.3,"Iris-virginica"); AddSample(117,6.5,3.0,5.5,1.8,"Iris-virginica"); AddSample(118,7.7,3.8,6.7,2.2,"Iris-virginica"); AddSample(119,7.7,2.6,6.9,2.3,"Iris-virginica"); AddSample(120,6.0,2.2,5.0,1.5,"Iris-virginica"); AddSample(121,6.9,3.2,5.7,2.3,"Iris-virginica"); AddSample(122,5.6,2.8,4.9,2.0,"Iris-virginica"); AddSample(123,7.7,2.8,6.7,2.0,"Iris-virginica"); AddSample(124,6.3,2.7,4.9,1.8,"Iris-virginica"); AddSample(125,6.7,3.3,5.7,2.1,"Iris-virginica"); AddSample(126,7.2,3.2,6.0,1.8,"Iris-virginica"); AddSample(127,6.2,2.8,4.8,1.8,"Iris-virginica"); AddSample(128,6.1,3.0,4.9,1.8,"Iris-virginica"); AddSample(129,6.4,2.8,5.6,2.1,"Iris-virginica"); AddSample(130,7.2,3.0,5.8,1.6,"Iris-virginica"); AddSample(131,7.4,2.8,6.1,1.9,"Iris-virginica"); AddSample(132,7.9,3.8,6.4,2.0,"Iris-virginica"); AddSample(133,6.4,2.8,5.6,2.2,"Iris-virginica"); AddSample(134,6.3,2.8,5.1,1.5,"Iris-virginica"); AddSample(135,6.1,2.6,5.6,1.4,"Iris-virginica"); AddSample(136,7.7,3.0,6.1,2.3,"Iris-virginica"); AddSample(137,6.3,3.4,5.6,2.4,"Iris-virginica"); AddSample(138,6.4,3.1,5.5,1.8,"Iris-virginica"); AddSample(139,6.0,3.0,4.8,1.8,"Iris-virginica"); AddSample(140,6.9,3.1,5.4,2.1,"Iris-virginica"); AddSample(141,6.7,3.1,5.6,2.4,"Iris-virginica"); AddSample(142,6.9,3.1,5.1,2.3,"Iris-virginica"); AddSample(143,5.8,2.7,5.1,1.9,"Iris-virginica"); AddSample(144,6.8,3.2,5.9,2.3,"Iris-virginica"); AddSample(145,6.7,3.3,5.7,2.5,"Iris-virginica"); AddSample(146,6.7,3.0,5.2,2.3,"Iris-virginica"); AddSample(147,6.3,2.5,5.0,1.9,"Iris-virginica"); AddSample(148,6.5,3.0,5.2,2.0,"Iris-virginica"); AddSample(149,6.2,3.4,5.4,2.3,"Iris-virginica"); AddSample(150,5.9,3.0,5.1,1.8,"Iris-virginica"); //--- ArrayResize(iris_samples,150); for(int i=0; i<Exttotal; i++) { iris_samples[i]=ExtIRISDataset[i]; } //--- return(true); } //+------------------------------------------------------------------+
分类方法说明：SVC、LinearSVC 和 NuSVC
让我们比较一下三种流行的分类方法：SVC（Support Vector Classification，支持向量分类）、LinearSVC（Linear Support Vector Classification，线性支持向量分类）和 NuSVC（Nu Support Vector Classification，核支持向量分类）。
操作原理：
SVC（Support Vector Classification）
工作准则：SVC 是一种基于最大化类别间边界的分类方法。它寻求一个最佳分离超平面，最大限度地分离类别并支持支持向量（最接近超平面的点）。
核函数：SVC 可以使用各种核函数，例如线性、径向基函数 (RBF)、多项式等。核函数决定如何转换数据以找到最佳超平面。
LinearSVC (Linear Support Vector Classification)
工作准则：LinearSVC 是 SVC 的一个变体，专门用于线性分类。它寻求一个不使用核函数的最佳线性分离超平面。这使得处理大量数据时更快、更高效。
NuSVC (Nu Support Vector Classification)
工作准则：NuSVC 也基于支持向量方法，但引入了一个参数 Nu (nu)，它控制模型的复杂性和支持向量的比例。Nu 值在 0 到 1 之间，决定了多少数据可用于支持向量和误差。
优点：
SVC
强大的算法：由于使用了核函数，SVC 可以处理复杂的分类任务并处理非线性数据。
对异常值的鲁棒性：SVC 对数据异常值具有很强的鲁棒性，因为它使用支持向量来构建分离超平面。
LinearSVC
高效率：LinearSVC 在处理大型数据集时更快、更高效，尤其是当数据量很大且线性分离适合该任务时。
线性分类:如果问题是线性可分的，LinearSVC 可以产生良好的结果，而不需要复杂的核函数。
NuSVC
模型复杂度控制：NuSVC 中的 Nu 参数允许您控制模型的复杂性以及拟合数据和泛化之间的权衡。
对异常值的鲁棒性：与 SVC 类似，NuSVC 对异常值具有很强的鲁棒性，这使得它对于处理噪声数据的任务非常有用。
限制：
SVC
计算复杂性：当处理大型数据集和/或使用复杂的核函数时，SVC 可能会很慢。
核敏感度：选择正确的核函数可能是一项具有挑战性的任务，并且会显著影响模型性能。
LinearSVC
线性约束：LinearSVC 受到线性数据分离的限制，在特征和目标变量之间存在非线性依赖关系的情况下表现不佳。
NuSVC
Nu 参数调整：调整 Nu 参数可能需要时间和实验才能获得最佳结果。
根据任务特点和数据量，每一种方法都可以是最佳选择。进行实验并选择最适合特定分类任务要求的方法非常重要。
让我们比较一下三种流行的分类方法：SVC（Support Vector Classification，支持向量分类）、LinearSVC（Linear Support Vector Classification，线性支持向量分类）和 NuSVC（Nu Support Vector Classification，核支持向量分类）。
操作原理：
SVC（Support Vector Classification）
工作准则：SVC 是一种基于最大化类别间边界的分类方法。它寻求一个最佳分离超平面，最大限度地分离类别并支持支持向量（最接近超平面的点）。
核函数：SVC 可以使用各种核函数，例如线性、径向基函数 (RBF)、多项式等。核函数决定如何转换数据以找到最佳超平面。
LinearSVC (Linear Support Vector Classification)
工作准则：LinearSVC 是 SVC 的一个变体，专门用于线性分类。它寻求一个不使用核函数的最佳线性分离超平面。这使得处理大量数据时更快、更高效。
NuSVC (Nu Support Vector Classification)
工作准则：NuSVC 也基于支持向量方法，但引入了一个参数 Nu (nu)，它控制模型的复杂性和支持向量的比例。Nu 值在 0 到 1 之间，决定了多少数据可用于支持向量和误差。
优点：
SVC
强大的算法：由于使用了核函数，SVC 可以处理复杂的分类任务并处理非线性数据。
对异常值的鲁棒性：SVC 对数据异常值具有很强的鲁棒性，因为它使用支持向量来构建分离超平面。
LinearSVC
高效率：LinearSVC 在处理大型数据集时更快、更高效，尤其是当数据量很大且线性分离适合该任务时。
线性分类:如果问题是线性可分的，LinearSVC 可以产生良好的结果，而不需要复杂的核函数。
NuSVC
模型复杂度控制：NuSVC 中的 Nu 参数允许您控制模型的复杂性以及拟合数据和泛化之间的权衡。
对异常值的鲁棒性：与 SVC 类似，NuSVC 对异常值具有很强的鲁棒性，这使得它对于处理噪声数据的任务非常有用。
限制：
SVC
计算复杂性：当处理大型数据集和/或使用复杂的核函数时，SVC 可能会很慢。
核敏感度：选择正确的核函数可能是一项具有挑战性的任务，并且会显著影响模型性能。
LinearSVC
线性约束：LinearSVC 受到线性数据分离的限制，在特征和目标变量之间存在非线性依赖关系的情况下表现不佳。
NuSVC
Nu 参数调整：调整 Nu 参数可能需要时间和实验才能获得最佳结果。
根据任务特点和数据量，每一种方法都可以是最佳选择。进行实验并选择最适合特定分类任务要求的方法非常重要。
2.1.SVC Classifier
Support Vector Classification（SVC，支持向量分类）分类方法是一种强大的机器学习算法，广泛用于解决分类任务。
操作原理：
- 最优分离超平面
工作准则：SVC 背后的主要思想是在特征空间中找到最佳分离超平面。该超平面应该最大化不同类别的对象之间的分离并且支持支持向量，这些支持向量是最接近超平面的数据点。
最大化间隔：SVC 的目标是最大化类之间的间隔，也就是支持向量到超平面的距离。这使得该方法对异常值具有鲁棒性，并能很好地推广到新数据。 - 核函数的使用
核函数：SVC 可以使用各种核函数，例如线性、径向基函数 (RBF)、多项式等。核函数允许将数据投影到更高维的空间中，即使在原始数据空间中没有线性可分性，任务也变得线性化。
核函数的选择：选择正确的核函数会显著影响 SVC 模型的性能。线性超平面并不总是最佳解决方案。
优点：
- 强大的算法。处理复杂任务：SVC 可以解决复杂的分类任务，包括特征和目标变量之间具有非线性依赖关系的任务。
- 对异常值的鲁棒性：支持向量的使用使得该方法对于数据异常值具有鲁棒性。它依赖于支持向量而不是整个数据集。
- 核函数的灵活性。数据适应性：使用不同核函数的能力使 SVC 能够适应特定数据并发现非线性关系。
- 很好的泛化能力。推广至新数据：SVC 模型可以很好地推广到新数据，使其可用于预测任务。
限制：
- 计算复杂性。训练时间：SVC 训练速度可能较慢，尤其是在处理大量数据或复杂核函数时。
- 核函数的选择。选择正确的核函数：选择正确的核函数可能需要实验，并且取决于数据特性。
- 对特征缩放的敏感性。数据规范化：SVC 对特征缩放很敏感，因此建议在训练之前对数据进行规范化或标准化。
- 模型可解释性。解释复杂性：由于使用非线性核函数和大量支持向量，SVC 模型的解释可能很复杂。
根据具体任务和数据量，SVC 方法可以成为解决分类任务的有力工具。然而，必须考虑其局限性并调整参数以获得最佳结果。
2.1.1.创建 SVC Classifier 模型的代码
此代码演示了在鸢尾花数据集上训练 SVC Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_SVCClassifier.py # The code demonstrates the process of training SVC model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.svm import SVC from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create an SVC Classifier model with a linear kernel svc_model = SVC(kernel='linear', C=1.0) # train the model on the entire dataset svc_model.fit(X, y) # predict classes for the entire dataset y_pred = svc_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of SVC Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(svc_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path +"svc_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of SVC Classifier model in ONNX format:", accuracy_onnx)
使用“编译”按钮在 MetaEditor 中运行脚本后，您可以在日志选项卡中查看其执行结果。
图 12.MetaEditor 中的 Iris_SVMClassifier.py 脚本的结果
Iris_SVCClassifier.py 脚本的输出：
Python Accuracy of SVC Classifier model:0.9933333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 0.98 0.99 50
Python 2 0.98 1.00 0.99 50
Python
Python accuracy 0.99 150
Python macro avg 0.99 0.99 0.99 150
Python weighted avg 0.99 0.99 0.99 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\svc_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of SVC Classifier model in ONNX format:0.9933333333333333
在这里，您可以找到有关 ONNX 模型保存的路径、ONNX 模型的输入和输出参数的类型以及描述鸢尾花数据集的准确性的信息。
使用 SVM 分类器描述数据集的准确率为 99%，导出为 ONNX 格式的模型也显示出相同的准确率。
现在，我们将通过对 150 个数据样本中的每一个运行构建的模型来在 MQL5 中验证这些结果。此外，该脚本还包括批量数据处理的示例。
2.1.2.使用 SVC Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_SVCClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "svc_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- ulong input_shape[]= { batch_size, input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- int output1[]; float output2[][3]; //--- ArrayResize(output1,(int)batch_size); ArrayResize(output2,(int)batch_size); //--- ulong output_shape[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {batch_size,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,ONNX_DEBUG_LOGS,input_data,output1,output2); //--- classes are ready in output1[k]; if(res) { for(int k=0; k<(int)batch_size; k++) model_classes_id[k]=output1[k]; } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="SVCClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
脚本执行的结果显示在 MetaTrader 5 终端的“专家”选项卡中。
Iris_SVCClassifier (EURUSD,H1) model:SVCClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_SVCClassifier (EURUSD,H1) model:SVCClassifier correct results: 99.33% Iris_SVCClassifier (EURUSD,H1) model=SVCClassifier all samples accuracy=0.993333 Iris_SVCClassifier (EURUSD,H1) model=SVCClassifier batch test accuracy=1.000000
SVC 模型对 150 个样本中的 149 个进行了正确分类，这是一个非常好的结果。该模型在鸢尾花数据集中仅出现一次分类错误，将样本 #84 预测为第 2 类（versicolor），而不是第 1 类（virginica）。
值得注意的是，导出的 ONNX 模型在完整鸢尾花数据集上的准确率为 99.33%，与原始模型的准确率相匹配。
2.1.3.SVC Classifier 模型的 ONNX 表示
您可以在 MetaEditor 中查看构建的 ONNX 模型。
图 13.MetaEditor 中的 ONNX 模型 svc_iris.onnx
有关模型架构的更多详细信息，您可以使用 Netron。为此，请单击 MetaEditor 中模型描述中的“在 Netron 中打开”按钮。
图 14.Netron 中的 ONNX 模型 svc_iris.onnx
图 15.Netron 中的 ONNX 模型 svc_iris.onnx（SVMClassifier ONNX 运算符参数）
2.2.LinearSVC Classifier
LinearSVC（Linear Support Vector Classification，线性支持向量分类）是一种强大的机器学习算法，用于二元和多类分类任务。它基于寻找最佳分离数据的超平面的想法。
LinearSVC 的原理：
- 寻找最优超平面：LinearSVC 的主要思想是找到最大限度分离两类数据的最佳超平面。超平面是由线性方程定义的多维平面。
- 最小化间隔：LinearSVC 旨在最小化间隔（数据点和超平面之间的距离）。间隔越大，超平面分离类别的效果就越好。
- 处理线性不可分数据：由于使用了核函数（核技巧），LinearSVC 可以处理原始特征空间中无法线性分离的数据，该核函数将数据投影到可以线性分离的高维空间中。
LinearSVC 的优点：
- 很好的泛化：LinearSVC 具有良好的泛化能力，在新的、未见过的数据上能表现良好。
- 效率：LinearSVC 能够快速处理大型数据集，并且需要的计算资源相对较少。
- 处理线性不可分数据：使用核函数，LinearSVC 可以解决具有线性不可分数据的分类任务。
- 可扩展性：LinearSVC 可以高效地用于具有大量特征和海量数据的任务。
LinearSVC 的局限性：
- 仅线性分离超平面：LinearSVC 仅构建线性分离超平面，这对于具有非线性依赖关系的复杂分类任务可能不够用。
- 参数选择：选择正确的参数（例如，正则化参数）可能需要专家知识或交叉验证。
- 对异常值的敏感性：LinearSVC 对数据中的异常值很敏感，这会影响分类质量。
- 模型可解释性：与其他一些方法相比，使用 LinearSVC 创建的模型可能不太容易解释。
LinearSVC 是一种强大的分类算法，具有出色的泛化能力、效率以及处理线性不可分数据的能力。它可用于各种分类任务，尤其是当数据可以通过线性超平面分离时。然而，对于需要建模非线性依赖关系的复杂任务，LinearSVC 可能不太合适，在这种情况下，应该考虑具有更复杂决策边界的替代方法。
2.2.1.创建 LinearSVC Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 LinearSVC Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_LinearSVC.py # The code demonstrates the process of training LinearSVC model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.svm import LinearSVC from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a LinearSVC model linear_svc_model = LinearSVC(C=1.0, max_iter=10000) # train the model on the entire dataset linear_svc_model.fit(X, y) # predict classes for the entire dataset y_pred = linear_svc_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of LinearSVC model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(linear_svc_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "linear_svc_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of LinearSVC model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of LinearSVC model:0.9666666666666667
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.96 0.94 0.95 50
Python 2 0.94 0.96 0.95 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\linear_svc_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of LinearSVC model in ONNX format:0.9666666666666667
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.96 0.94 0.95 50
Python 2 0.94 0.96 0.95 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\linear_svc_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of LinearSVC model in ONNX format:0.9666666666666667
2.2.2.用于处理 LinearSVC Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_LinearSVC.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "linear_svc_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- ulong input_shape[]= { batch_size, input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- int output1[]; float output2[][3]; //--- ArrayResize(output1,(int)batch_size); ArrayResize(output2,(int)batch_size); //--- ulong output_shape[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {batch_size,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,ONNX_DEBUG_LOGS,input_data,output1,output2); //--- classes are ready in output1[k]; if(res) { for(int k=0; k<(int)batch_size; k++) model_classes_id[k]=output1[k]; } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="LinearSVC"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_LinearSVC (EURUSD,H1) model:LinearSVC sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_LinearSVC (EURUSD,H1) model:LinearSVC sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_LinearSVC (EURUSD,H1) model:LinearSVC sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_LinearSVC (EURUSD,H1) model:LinearSVC sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_LinearSVC (EURUSD,H1) model:LinearSVC sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_LinearSVC (EURUSD,H1) model:LinearSVC correct results: 96.67% Iris_LinearSVC (EURUSD,H1) model=LinearSVC all samples accuracy=0.966667 Iris_LinearSVC (EURUSD,H1) model=LinearSVC batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 96.67%，与原始模型的准确率一致。
2.2.3.LinearSVC Classifier 模型的 ONNX 表示
图 16.Netron 中 LinearSVC Classifier 模型的 ONNX 表示
2.3.NuSVC Classifier
Nu-Support Vector Classification (NuSVC，核支持向量分类 ) 方法是一种基于支持向量机 (SVM) 方法的强大的机器学习算法。
NuSVC 的原理：
- 支持向量机（SVM）：NuSVC 是 SVM 的变体，用于二元和多类分类任务。SVM的核心原理就是寻找最优分离超平面，在保持最大间隔的同时，最大限度地分离类别。
- Nu 参数：NuSVC 中的一个关键参数是 Nu 参数（nu），它控制模型的复杂性并定义可用作支持向量和误差的样本的比例。Nu 的值范围是 0 到 1，其中 0.5 表示大约一半的样本将被用作支持向量和误差。
- 参数调整：确定 Nu 参数和其他超参数的最佳值可能需要交叉验证并在训练数据中搜索最佳值。
- 核函数：NuSVC 可以使用各种核函数，例如线性、径向基函数 (RBF)、多项式等。核函数决定如何变换特征空间以找到分离超平面。
NuSVC 的优点：
- 高维空间中的效率：NuSVC 可以在高维空间中高效工作，适合具有大量特征的任务。
- 对异常值的鲁棒性：由于使用了支持向量，SVM，尤其是 NuSVC，对数据中的异常值具有很强的鲁棒性。
- 模型复杂度的控制：Nu 参数可以控制模型复杂性并平衡数据拟合和泛化。
- 良好的泛化：尤其是 SVM 和 NuSVC，表现出了良好的泛化能力，在新的、以前从未见过的数据上有着优异的表现。
NuSVC 的局限性：
- 大数据量导致效率低下：由于计算复杂性，NuSVC 在对大量数据进行训练时效率低下。
- 所需参数调整：调整 Nu 参数和核函数可能需要时间和计算资源。
- 核函数线性：NuSVC 的有效性在很大程度上取决于核函数的选择，对于某些任务，可能需要尝试不同的函数。
- 模型可解释性：SVM 和 NuSVC 提供了出色的结果，但它们的模型解释起来可能很复杂，尤其是在使用非线性核函数时。
Nu-Support Vector Classification（NuSVC）是一种基于 SVM 的强大分类方法，具有对异常值的鲁棒性和良好的泛化性等优点。然而，它的有效性取决于参数和核函数的选择，并且对于大量数据来说可能效率低下。仔细选择参数并使方法适应特定的分类任务至关重要。
2.3.1.创建 NuSVC Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 NuSVC Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_NuSVC.py # The code demonstrates the process of training NuSVC model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.svm import NuSVC from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a NuSVC model nusvc_model = NuSVC(nu=0.5, kernel='linear') # train the model on the entire dataset nusvc_model.fit(X, y) # predict classes for the entire dataset y_pred = nusvc_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of NuSVC model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(nusvc_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "nusvc_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of NuSVC model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of NuSVC model:0.9733333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.96 0.96 0.96 50
Python 2 0.96 0.96 0.96 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\nusvc_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of NuSVC model in ONNX format:0.9733333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.96 0.96 0.96 50
Python 2 0.96 0.96 0.96 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\nusvc_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of NuSVC model in ONNX format:0.9733333333333334
2.3.2.用于处理 NuSVC Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_NuSVC.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "nusvc_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- ulong input_shape[]= { batch_size, input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- int output1[]; float output2[][3]; //--- ArrayResize(output1,(int)batch_size); ArrayResize(output2,(int)batch_size); //--- ulong output_shape[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {batch_size,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,ONNX_DEBUG_LOGS,input_data,output1,output2); //--- classes are ready in output1[k]; if(res) { for(int k=0; k<(int)batch_size; k++) model_classes_id[k]=output1[k]; } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="NuSVC"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_NuSVC (EURUSD,H1) model:NuSVC sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_NuSVC (EURUSD,H1) model:NuSVC sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_NuSVC (EURUSD,H1) model:NuSVC sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_NuSVC (EURUSD,H1) model:NuSVC sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_NuSVC (EURUSD,H1) model:NuSVC correct results: 97.33% Iris_NuSVC (EURUSD,H1) model=NuSVC all samples accuracy=0.973333 Iris_NuSVC (EURUSD,H1) model=NuSVC batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 97.33%，与原始模型的准确率一致。
2.3.3.NuSVC Classifier 模型的 ONNX 表示
图 17.Netron 中 NuSVC Classifier 模型的 ONNX 表示
2.4.Radius Neighbors Classifier
Radius Neighbors Classifier（半径邻域分类器）是一种基于对象之间接近度原理进行分类任务的机器学习方法。与经典的 K-Nearest Neighbors (K-NN) Classifier 不同，在 K-NN Classifier 中选择的是固定数量的最近邻，而在 Radius Neighbors Classifier中，对象是根据与指定半径内的最近邻的距离进行分类的。Radius Neighbors Classifier 的原理：
- 确定半径：Radius Neighbors Classifier 的主要参数是半径，它定义了一个对象与其邻居之间的最大距离，以使其被认为接近邻居的类别。
- 寻找最近邻居：计算每个对象与训练数据集中所有其他对象的距离。位于指定半径内的物体被视为该物体的邻居。
- 投票：Radius Neighbors Classifier 使用邻居之间的多数投票来确定对象的类别。例如，如果大多数邻居属于 A 类，则该对象也将被归类为 A 类。
- 对数据密度的适应性：Radius Neighbors Classifier 适用于不同特征空间区域中的数据密度可能变化的任务。
- 能够与不同的类别形状一起工作：该方法在类别具有复杂和非线性形状的任务中表现良好。
- 适用于有异常值的数据：Radius Neighbors Classifier 对异常值的鲁棒性比 K-NN 更强，因为它忽略了位于指定半径之外的邻居。
- 对半径选择的敏感性：选择最佳半径值并非易事，需要进行调整。
- 大型数据集上的效率低下：对于大型数据集，计算到所有对象的距离可能会非常耗费计算资源。
- 对数据密度的依赖：当数据在特征空间中密度不均匀时，该方法可能效率较低。
在对象接近度很重要且类别形状可能很复杂的情况下，Radius Neighbors Classifier 是一种很有价值的机器学习方法。它可以应用于各个领域，包括图像分析、自然语言处理等。
2.4.1.创建 Radius Neighbors Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Radius Neighbors Classifier 模型、以 ONNX 格式导出以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_RadiusNeighborsClassifier.py # The code demonstrates the process of training an Radius Neughbors model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023 MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.neighbors import RadiusNeighborsClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Radius Neighbors Classifier model radius_model = RadiusNeighborsClassifier(radius=1.0) # train the model on the entire dataset radius_model.fit(X, y) # predict classes for the entire dataset y_pred = radius_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Radius Neighbors Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(radius_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "radius_neighbors_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Radius Neighbors Classifier model in ONNX format:", accuracy_onnx)
脚本 Iris_RadiusNeighbors.py 的结果：
Python Accuracy of Radius Neighbors Classifier model:0.9733333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.98 0.96 50
Python 2 0.98 0.94 0.96 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\radius_neighbors_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of Radius Neighbors Classifier model in ONNX format:0.9733333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.98 0.96 50
Python 2 0.98 0.94 0.96 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\radius_neighbors_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of Radius Neighbors Classifier model in ONNX format:0.9733333333333334
原始模型的精度与以ONNX格式导出的模型的精度相同。
2.4.2.用于处理 Radius Neighbors Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_RadiusNeighborsClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "radius_neighbors_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- ulong input_shape[]= { batch_size, input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- int output1[]; float output2[][3]; //--- ArrayResize(output1,(int)batch_size); ArrayResize(output2,(int)batch_size); //--- ulong output_shape[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {batch_size,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,ONNX_DEBUG_LOGS,input_data,output1,output2); //--- classes are ready in output1[k]; if(res) { for(int k=0; k<(int)batch_size; k++) model_classes_id[k]=output1[k]; } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="RadiusNeighborsClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_RadiusNeighborsClassifier (EURUSD,H1) model:RadiusNeighborsClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_RadiusNeighborsClassifier (EURUSD,H1) model:RadiusNeighborsClassifier sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_RadiusNeighborsClassifier (EURUSD,H1) model:RadiusNeighborsClassifier sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_RadiusNeighborsClassifier (EURUSD,H1) model:RadiusNeighborsClassifier sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_RadiusNeighborsClassifier (EURUSD,H1) model:RadiusNeighborsClassifier correct results: 97.33% Iris_RadiusNeighborsClassifier (EURUSD,H1) model=RadiusNeighborsClassifier all samples accuracy=0.973333 Iris_RadiusNeighborsClassifier (EURUSD,H1) model=RadiusNeighborsClassifier batch test accuracy=1.000000
Radius Neighbors Classifier 模型的准确率为 97.33%，分类错误数为 4 个（样本 78、107、127 和 139）。
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 97.33%，与原始模型的准确率相匹配。
2.4.3. Radius Neighbors Classifier 模型的 ONNX 表示
图 18.Netron 中 Radius Neighbors Classifier 的 ONNX 表示
关于 RidgeClassifier 和 RidgeClassifierCV 方法的注意事项
RidgeClassifier 和 RidgeClassifierCV 是两种基于 Ridge Regression（岭回归）的分类方法，但它们在参数调整和超参数自动选择的方式上有所不同：
RidgeClassifier:
- RidgeClassifier 是一种基于 Ridge Regression 的分类方法，用于二元和多类分类任务。
- 在多类分类的情况下，RidgeClassifier 将任务转换为多个二元任务（一对多）并为每个任务建立一个模型。
- 正则化参数 alpha 需要用户手动调整，这意味着您必须通过实验或分析验证数据来选择最佳的 alpha 值。
RidgeClassifierCV:
- RidgeClassifierCV 是 RidgeClassifier 的扩展，它提供对交叉验证和自动选择最佳正则化参数 alpha 的内置支持。
- 您无需手动设置 alpha，而是可以向 RidgeClassifierCV 提供一个 alpha 值列表来调查并指定交叉验证方法（例如，通过 cv 参数）。
- RidgeClassifierCV 自动选择在交叉验证期间表现最优的最佳 alpha 值。
因此，它们之间的主要区别在于选择正则化参数 alpha 的最佳值的自动化程度。RidgeClassifier 需要手动调整 alpha，而 RidgeClassifierCV 允许使用交叉验证自动选择最佳 alpha 值。它们之间的选择取决于您的需求以及您对模型调整过程自动化的渴望。
2.5.Ridge Classifier
Ridge Classifier 是逻辑回归的一种变体，在模型中包含了 L2 正则化（岭回归）。L2正则化对模型的较大系数添加了惩罚，有助于减少过拟合，提高模型的泛化能力。
Ridge Classifier 的原理：
- 概率预测：与逻辑回归类似，Ridge Classifier 使用逻辑（sigmoid）函数对对象属于特定类别的概率进行建模。
- L2 正则化：Ridge Classifier 添加了一个 L2 正则化项，用于惩罚模型的较大系数。这样做是为了控制模型的复杂性并减少过度拟合。
- 参数训练：在训练数据集上训练 Ridge Classifier 模型，以调整特征的权重（系数）和正则化参数。
Ridge Classifier 的优点：
- 减少过度拟合：L2 正则化有助于降低模型过度拟合的趋势，这在数据有限时尤其有用。
- 处理多重共线性（Multicollinearity）：Ridge Classifier 可以很好地处理多重共线性，即其中特征彼此高度相关的问题。
Ridge Classifier 的局限性：
- 对正则化参数选择的敏感性：与其他正则化方法一样，选择正确的正则化参数（alpha）值需要调整和评估。
- 多类分类约束：Ridge Classifier 最初设计用于二元分类，但可以使用 One-vs-All （一对多）等方法适应多类分类。
Ridge Classifier 是一种强大的机器学习方法，它结合了逻辑回归与正则化的优点，以对抗过度拟合并提高模型的泛化能力。它在概率分类和模型复杂性控制很重要的各个领域都有应用。
2.5.1.Ridge Classifier 模型创建代码
此代码演示了在 Iris 数据集上训练 Ridge Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_RidgeClassifier.py # The code demonstrates the process of training Ridge Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import RidgeClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Ridge Classifier model ridge_model = RidgeClassifier() # train the model on the entire dataset ridge_model.fit(X, y) # predict classes for the entire dataset y_pred = ridge_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Ridge Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(ridge_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "ridge_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Ridge Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Ridge Classifier model:0.8533333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.87 0.66 0.75 50
Python 2 0.73 0.90 0.80 50
Python
Python accuracy 0.85 150
Python macro avg 0.86 0.85 0.85 150
Python weighted avg 0.86 0.85 0.85 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\ridge_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of Ridge Classifier model in ONNX format:0.8533333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.87 0.66 0.75 50
Python 2 0.73 0.90 0.80 50
Python
Python accuracy 0.85 150
Python macro avg 0.86 0.85 0.85 150
Python weighted avg 0.86 0.85 0.85 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\ridge_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of Ridge Classifier model in ONNX format:0.8533333333333334
2.5.2.用于处理 Ridge Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_RidgeClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "ridge_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- ulong input_shape[]= { batch_size, input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- int output1[]; float output2[][3]; //--- ArrayResize(output1,(int)batch_size); ArrayResize(output2,(int)batch_size); //--- ulong output_shape[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {batch_size,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,ONNX_DEBUG_LOGS,input_data,output1,output2); //--- classes are ready in output1[k]; if(res) { for(int k=0; k<(int)batch_size; k++) model_classes_id[k]=output1[k]; } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="RidgeClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=51 FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=52 FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=57 FAILED [class=2, true class=1] features=(6.30,3.30,4.70,1.60] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=62 FAILED [class=2, true class=1] features=(5.90,3.00,4.20,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=65 FAILED [class=2, true class=1] features=(5.60,2.90,3.60,1.30] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=66 FAILED [class=2, true class=1] features=(6.70,3.10,4.40,1.40] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=76 FAILED [class=2, true class=1] features=(6.60,3.00,4.40,1.40] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=79 FAILED [class=2, true class=1] features=(6.00,2.90,4.50,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=86 FAILED [class=2, true class=1] features=(6.00,3.40,4.50,1.60] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=87 FAILED [class=2, true class=1] features=(6.70,3.10,4.70,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=89 FAILED [class=2, true class=1] features=(5.60,3.00,4.10,1.30] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=92 FAILED [class=2, true class=1] features=(6.10,3.00,4.60,1.40] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=109 FAILED [class=1, true class=2] features=(6.70,2.50,5.80,1.80] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier correct results: 85.33% Iris_RidgeClassifier (EURUSD,H1) model=RidgeClassifier all samples accuracy=0.853333 Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40) Iris_RidgeClassifier (EURUSD,H1) model:RidgeClassifier FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50) Iris_RidgeClassifier (EURUSD,H1) model=RidgeClassifier batch test accuracy=0.000000
在完整的 Iris 数据集上，该模型表现出 85.33% 的准确率，与原始准确率一致。
2.5.3.Ridge Classifier 模型的 ONNX 表示
图 19.Netron 中 Ridge Classifier 模型的 ONNX 表示
2.6.RidgeClassifierCV
RidgeClassifierCV 分类方法是一种基于岭回归的二元和多类分类的强大算法。
RidgeClassifierCV 的原理：
- 线性岭回归：RidgeClassifierCV 基于线性岭回归。该方法是对线性回归的修改，其中添加了 L2 正则化。正则化通过降低特征权重的大小来帮助控制过度拟合。
- 二元和多类分类：RidgeClassifierCV 既可用于二分类（只有两个类时），也可用于多类分类（有两个以上的类时）。对于多类分类，它将任务转换为多个二元任务（一对多）并为每个任务建立一个模型。
- 正则化参数的自动选择：RidgeClassifierCV 的一个主要优势是它内置对交叉验证和自动选择最佳正则化参数 alpha 的支持。该方法不需要手动调整 alpha，而是迭代不同的 alpha 值并根据交叉验证选择最佳值。
- 处理多重共线性（Multicollinearity）：岭回归可以很好地处理多重共线性问题，其中特征彼此高度相关。正则化可以控制每个特征的贡献，使得模型对相关数据具有鲁棒性。
RidgeClassifierCV 的优点：
- 自动超参数选择：RidgeClassifierCV 的一个显著优势是它能够使用交叉验证自动选择最佳 alpha 值。这消除了尝试不同 alpha 值的需要，并增加了获得良好结果的可能性。
- 过度拟合控制：RidgeClassifierCV 提供的 L2 正则化有助于控制模型复杂性并降低过度拟合的风险。这对于数据有限的任务尤其重要。
- 透明度和可解释性：RidgeClassifierCV 提供可解释的特征权重，允许分析每个特征对预测的贡献并得出特征重要性结论。
- 效率：该方法效率高，可应用于大型数据集。
RidgeClassifierCV 的局限性：
- 线性：RidgeClassifierCV 假设特征和目标变量之间存在线性关系。如果数据表现出强烈的非线性关系，该方法可能不够准确。
- 特征缩放敏感度：该方法对特征缩放很敏感。建议在应用 RidgeClassifierCV 之前对特征进行标准化或规范化。
- 最佳特征选择：RidgeClassifierCV 不执行自动特征选择，因此您需要手动决定在模型中包含哪些特征。
RidgeClassifierCV 分类方法是一个强大的二元和多类分类工具，可以自动选择最佳正则化参数。它的过度拟合控制、可解释性和效率使其成为各种分类任务的热门选择。然而，重要的是要记住它的局限性，特别是特征和目标变量之间的线性关系的假设。
2.6.1.RidgeClassifierCV 模型创建代码
此代码演示了在 Iris 数据集上训练 RidgeClassifierCV 模型、导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_RidgeClassifierCV.py # The code demonstrates the process of training RidgeClassifierCV model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import RidgeClassifierCV from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a RidgeClassifierCV model ridge_classifier_cv_model = RidgeClassifierCV() # train the model on the entire dataset ridge_classifier_cv_model.fit(X, y) # predict classes for the entire dataset y_pred = ridge_classifier_cv_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of RidgeClassifierCV model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(ridge_classifier_cv_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "ridge_classifier_cv_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of RidgeClassifierCV model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of RidgeClassifierCV model:0.8533333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.87 0.66 0.75 50
Python 2 0.73 0.90 0.80 50
Python
Python accuracy 0.85 150
Python macro avg 0.86 0.85 0.85 150
Python weighted avg 0.86 0.85 0.85 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\ridge_classifier_cv_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of RidgeClassifierCV model in ONNX format:0.8533333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.87 0.66 0.75 50
Python 2 0.73 0.90 0.80 50
Python
Python accuracy 0.85 150
Python macro avg 0.86 0.85 0.85 150
Python weighted avg 0.86 0.85 0.85 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\ridge_classifier_cv_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: probabilities, Data Type: tensor(float), Shape: [None, 3]
Python
Python Accuracy of RidgeClassifierCV model in ONNX format:0.8533333333333334
2.6.2.用于处理 RidgeClassifierCV 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_RidgeClassifierCV.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "ridge_classifier_cv_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- ulong input_shape[]= { batch_size, input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- int output1[]; float output2[][3]; //--- ArrayResize(output1,(int)batch_size); ArrayResize(output2,(int)batch_size); //--- ulong output_shape[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {batch_size,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,ONNX_DEBUG_LOGS,input_data,output1,output2); //--- classes are ready in output1[k]; if(res) { for(int k=0; k<(int)batch_size; k++) model_classes_id[k]=output1[k]; } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="RidgeClassifierCV"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=51 FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=52 FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=57 FAILED [class=2, true class=1] features=(6.30,3.30,4.70,1.60] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=62 FAILED [class=2, true class=1] features=(5.90,3.00,4.20,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=65 FAILED [class=2, true class=1] features=(5.60,2.90,3.60,1.30] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=66 FAILED [class=2, true class=1] features=(6.70,3.10,4.40,1.40] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=76 FAILED [class=2, true class=1] features=(6.60,3.00,4.40,1.40] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=79 FAILED [class=2, true class=1] features=(6.00,2.90,4.50,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=86 FAILED [class=2, true class=1] features=(6.00,3.40,4.50,1.60] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=87 FAILED [class=2, true class=1] features=(6.70,3.10,4.70,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=89 FAILED [class=2, true class=1] features=(5.60,3.00,4.10,1.30] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=92 FAILED [class=2, true class=1] features=(6.10,3.00,4.60,1.40] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=109 FAILED [class=1, true class=2] features=(6.70,2.50,5.80,1.80] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV correct results: 85.33% Iris_RidgeClassifierCV (EURUSD,H1) model=RidgeClassifierCV all samples accuracy=0.853333 Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40) Iris_RidgeClassifierCV (EURUSD,H1) model:RidgeClassifierCV FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50) Iris_RidgeClassifierCV (EURUSD,H1) model=RidgeClassifierCV batch test accuracy=0.000000
ONNX 模型的性能也与原始 scikit-learn 模型的性能（85.33%）完美匹配。
2.6.3.RidgeClassifierCV 模型的 ONNX 表示
图 20.Netron 中 RidgeClassifierCV 的 ONNX 表示
2.7.Random Forest Classifier
Random Forest Classifier（随机森林分类器）是一种基于构建多棵决策树并结合其结果来提高分类质量的集成机器学习方法。该方法由于其有效性和处理各种数据的能力而非常受欢迎。
Random Forest Classifier 的原理：
- Bagging（Bootstrap Aggregating，自助聚集）：随机森林使用装袋方法，该方法涉及从训练数据中创建多个子样本（自助样本）并进行替换。对于每个子样本，都会构建一个单独的决策树。
- 随机特征选择：在构建每棵树时，会从整个特征集中选择一个随机的特征子集。这促进了树木之间的多样性并减少了它们之间的相关性。
- 投票：在对对象进行分类时，每棵树都提供自己的预测，并选择所有树中获得多数投票的类作为最终的模型预测。
Random Forest Classifier 的优点：
- 高准确率：随机森林通常通过对多棵树的结果进行平均来实现较高的分类准确率。
- 处理多样化数据的能力：它可以很好地处理数字和分类特征以及不同结构的数据。
- 抗过度拟合：随机森林具有内置正则化，使其能够抵抗过度拟合。
- 特征重要性：随机森林可以评估特征重要性，帮助数据科学家和特征工程师更好地理解数据。
Random Forest Classifier 的局限性：
- 计算复杂性：训练随机森林可能非常耗时，尤其是当有大量树木和特征时。
- 可解释性的挑战：由于树木数量众多且特征选择随机，模型解释可能具有挑战性。
- 无法保证异常值稳健性：随机森林并不总是能对数据异常值提供稳健性。
Random Forest Classifier 是一种强大的机器学习算法，广泛应用于生物医药、金融分析和文本数据分析等各个领域。它擅长解决分类和回归任务，具有很强的泛化能力。
2.7.1.Random Forest Classifier 模型创建代码
此代码演示了在 Iris 数据集上训练 Random Forest Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_RandomForestClassifier.py # The code demonstrates the process of training Random Forest Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023,2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.ensemble import RandomForestClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Random Forest Classifier model rf_model = RandomForestClassifier(n_estimators=100, random_state=42) # train the model on the entire dataset rf_model.fit(X, y) # predict classes for the entire dataset y_pred = rf_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Random Forest Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(rf_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "rf_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Random Forest Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Random Forest Classifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\rf_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Random Forest Classifier model in ONNX format:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\rf_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Random Forest Classifier model in ONNX format:1.0
Random Forest Classifier 模型（及其 ONNX 版本）以 100% 的准确率解决了 Fisher 的鸢尾花分类问题。
2.7.2.使用 Random Forest Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_RandomForestClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "rf_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="RandomForestClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_RandomForestClassifier (EURUSD,H1) model:RandomForestClassifier correct results: 100.00% Iris_RandomForestClassifier (EURUSD,H1) model=RandomForestClassifier all samples accuracy=1.000000 Iris_RandomForestClassifier (EURUSD,H1) model=RandomForestClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率相匹配。
2.7.3.Random Forest Classifier 模型的 ONNX 表示
图 21.Netron 中 Random Forest Classifier 模型的 ONNX 表示
2.8.Gradient Boosting Classifier
Gradient boosting（梯度提升）是最强大的机器学习方法之一，由于其高精度和处理多样化数据的能力，可应用于数据分析、计算机视觉、自然语言处理和金融分析等各个领域。Gradient Boosting Classifier 是一种集成机器学习方法，它建立决策树（decision trees）组合来解决分类任务。该方法因其能够实现高精度并且抗过度拟合而广受欢迎。
Gradient Boosting Classifier 的原理：
- 决策树组合：Gradient Boosting Classifier 构建了一组决策树，其中每棵树都旨在改善前一棵树的预测。
- 梯度下降：梯度提升使用梯度下降来优化损失函数。它通过计算损失函数的梯度并根据该梯度更新预测来最小化分类误差。
- 树权重：组合中的每棵树都有一个权重，最后，根据权重将所有树的预测组合起来。
Gradient Boosting Classifier 的优点：
- 高准确率：Gradient Boosting Classifier 通常提供较高的分类精度，是最强大的机器学习方法之一。
- 抗过度拟合：由于使用了正则化和梯度下降，该方法具有抗过度拟合的能力，尤其是在调整超参数时。
- 能够处理不同类型的数据类型：Gradient Boosting Classifier 可以处理各种数据类型，包括数字和分类特征。
Gradient Boosting Classifier 的局限性：
- 计算复杂性：训练 Gradient Boosting Classifier 的计算量很大，尤其是在存在大量树或深树的情况下。
- 可解释性的挑战：由于多棵树的组成非常复杂，解释结果可能具有挑战性。
- 并不总是适合小型数据集：梯度提升通常需要大量数据才能有效运行，并且在小数据集上容易过度拟合。
Gradient Boosting Classifier 是一种强大的机器学习方法，常用于数据分析竞赛，可以有效地解决各种分类任务。它可以发现数据中复杂的非线性关系，并且在超参数适当调整时表现出良好的泛化能力。
2.8.1.Gradient Boosting Classifier 模型创建代码
此代码演示了在 Iris 数据集上训练 Gradient Boosting Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_GradientBoostingClassifier.py # The code demonstrates the process of training Gradient Boostring Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.ensemble import GradientBoostingClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Gradient Boosting Classifier model gb_model = GradientBoostingClassifier(n_estimators=100, random_state=42) # train the model on the entire dataset gb_model.fit(X, y) # predict classes for the entire dataset y_pred = gb_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Gradient Boosting Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(gb_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "gb_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Gradient Boosting Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Gradient Boosting Classifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\gb_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Gradient Boosting Classifier model in ONNX format:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\gb_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Gradient Boosting Classifier model in ONNX format:1.0
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率相匹配。
2.8.2.用于处理 Gradient Boosting Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_GradientBoostingClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "gb_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="GradientBoostingClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_GradientBoostingClassifier (EURUSD,H1) model:GradientBoostingClassifier correct results: 100.00% Iris_GradientBoostingClassifier (EURUSD,H1) model=GradientBoostingClassifier all samples accuracy=1.000000 Iris_GradientBoostingClassifier (EURUSD,H1) model=GradientBoostingClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率相匹配。
2.8.3.Gradient Boosting Classifier 模型的 ONNX 表示
图 22.Netron 中 Gradient Boosting Classifier 模型的 ONNX 表示
2.9.Adaptive Boosting Classifier
AdaBoost（Adaptive Boosting，自适应增强）Classifier 是一种集成机器学习方法，通过组合多个弱（例如决策树）分类器的结果来创建更强大的算法，从而增强分类能力。
AdaBoost Classifier 的原理：
- 弱分类器集成：AdaBoost 首先初始化训练集中的每个样本的权重，并为它们分配相等的初始值。
- 训练弱分类器：然后，AdaBoost 会根据样本权重在训练集上训练弱分类器（例如决策树）。该分类器尝试正确地对样本进行分类。
- 权重重新分配：AdaBoost 通过调整样本权重，增加分类错误的样本的权重，减少分类正确的样本的权重。
- 创建组合：AdaBoost 重复训练弱分类器和重新分配权重的过程多次，然后将这些弱分类器的结果组合成一个组合，每个分类器根据其准确性做出贡献。
AdaBoost Classifier 的优点：
- 高准确率：AdaBoost 通常通过组合几个弱分类器来提供较高的分类准确率。
- 抗过度拟合：AdaBoost 具有内置正则化，使其能够抵抗过度拟合。
- 能够与各种分类器配合使用：AdaBoost 可以使用不同的基础分类器，从而适应特定的任务。
AdaBoost Classifier 的局限性：
- 对异常值的敏感性：AdaBoost 对数据中的异常值很敏感，因为它们可能具有很大的权重。
- 并不总是适合复杂的任务：在一些复杂的任务中，AdaBoost 可能需要大量的基础分类器才能取得良好的效果。
- 对基础分类器质量的依赖：当基础分类器优于随机猜测时，AdaBoost 的表现会更好。
AdaBoost Classifier 是一种强大的机器学习算法，在实践中常用于解决分类任务。它非常适合二元和多类问题，并且可以适应各种基础分类器。
2.9.1.AdaBoost Classifier 模型创建代码
此代码演示了在 Iris 数据集上训练 AdaBoost Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_AdaBoostClassifier.py # The code demonstrates the process of training AdaBoost Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.ensemble import AdaBoostClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create an AdaBoost Classifier model adaboost_model = AdaBoostClassifier(n_estimators=50, random_state=42) # train the model on the entire dataset adaboost_model.fit(X, y) # predict classes for the entire dataset y_pred = adaboost_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of AdaBoost Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(adaboost_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "adaboost_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of AdaBoost Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of AdaBoost Classifier model:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.92 0.96 0.94 50
Python 2 0.96 0.92 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\adaboost_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of AdaBoost Classifier model in ONNX format:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.92 0.96 0.94 50
Python 2 0.96 0.92 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\adaboost_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of AdaBoost Classifier model in ONNX format:0.96
2.9.2.用于处理 Adaptive Boosting Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_AdaBoostClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "adaboost_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="AdaBoostClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AdaBoostClassifier (EURUSD,H1) model:AdaBoostClassifier correct results: 96.00% Iris_AdaBoostClassifier (EURUSD,H1) model=AdaBoostClassifier all samples accuracy=0.960000 Iris_AdaBoostClassifier (EURUSD,H1) model=AdaBoostClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 96%，与原始模型的准确率一致。
2.9.3.AdaBoost Classifier 模型的 ONNX 表示
图 23.Netron 中 AdaBoost Classifier 的 ONNX 表示
2.10.Bootstrap Aggregating Classifier
Bagging（Bootstrap Aggregating，自助聚集）Classifier 是一种集成机器学习方法，它基于从训练数据中创建多个随机子样本（引导样本）并在每个子样本上构建单独的模型。然后将结果结合起来以提高模型的泛化能力。
Bagging Classifier 的原理：
- 创建子样本：Bagging 首先从训练数据中创建几个随机子样本（自助样本），并进行替换。这意味着相同的样本可能出现在多个子样本中，并且一些样本可能会被省略。
- 训练基础模型：在每个子样本上，训练一个单独的基础模型（例如，决策树）。每个模型都独立于其他模型进行训练。
- 结果汇总：训练完所有基础模型后，将它们的预测结果结合起来，得到最终的预测。在二元分类中，这可以通过多数投票来完成。
Bagging Classifier 的优点：
- 减少方差：Bagging 通过对多个基础模型的结果取平均值来减少模型的方差，从而可以得到更稳定、更可靠的预测。
- 减少过度拟合：由于每个基础模型都在不同的子样本上进行训练，Bagging 可以降低模型过度拟合的趋势。
- 多功能性：Bagging 可以使用各种基础模型，从而适应不同的数据类型和任务。
Bagging Classifier 的局限性：
- 不会改善偏差：Bagging 倾向于减少方差，但不能解决模型的偏差。如果基础模型倾向于出现偏差（例如，拟合不足），Bagging 将无法纠正这个问题。
- 并不总是适合复杂的任务：在一些复杂的任务中，Bagging可能需要大量的基础模型才能取得良好的效果。
Bagging Classifier 是一种有效的机器学习方法，可以增强模型的泛化能力，减少过度拟合。它通常与不同的基础模型结合使用来解决各种分类和回归任务。
2.10.1.Bootstrap Aggregating Classifier 模型创建代码
此代码演示了在 Iris 数据集上训练 Bootstrap Aggregating Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_BootstrapAggregatingClassifier.py # The code demonstrates the process of training Bagging Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.ensemble import BaggingClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Bagging Classifier model with a Decision Tree base estimator bagging_model = BaggingClassifier(n_estimators=100, random_state=42) # train the model on the entire dataset bagging_model.fit(X, y) # predict classes for the entire dataset y_pred = bagging_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Bagging Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(bagging_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "bagging_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Bagging Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Bagging Classifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\bagging_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Bagging Classifier model in ONNX format:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\bagging_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Bagging Classifier model in ONNX format:1.0
Bootstrap Aggregating Classifier 模型（及其 ONNX 版本）在对 Iris 数据集进行分类时实现了 100% 的准确率。
2.10.2.用于处理 Bootstrap Aggregating Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_BootstrapAggregatingClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "bagging_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="BootstrapAggregatingClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_BootstrapAggregatingClassifier (EURUSD,H1) model:BootstrapAggregatingClassifier correct results: 100.00% Iris_BootstrapAggregatingClassifier (EURUSD,H1) model=BootstrapAggregatingClassifier all samples accuracy=1.000000 Iris_BootstrapAggregatingClassifier (EURUSD,H1) model=BootstrapAggregatingClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为100%，与原始模型的准确率一致。
2.10.3.Bootstrap Aggregating Classifier 的 ONNX 表示
图 24.Netron 中 Bootstrap Aggregating Classifier 的 ONNX 表示
2.11.K-Nearest Neighbors (K-NN) Classifier
K-Nearest Neighbors (K-NN) 分类器是一种机器学习方法，用于根据数据点之间的相似性解决分类和回归任务。它的工作原理是，多维特征空间中彼此接近的对象具有相似的特征，因此可能具有相似的类别标签。
K-NN Classifier 的原理：
- 确定接近度：K-NN Classifier 计算待分类对象与训练数据集中其他对象之间的接近度。这通常使用距离度量来完成，例如欧几里得距离或曼哈顿距离。
- 选择邻居的数量：参数 K 决定了用于对对象进行分类的最近邻居的数量。通常，根据任务和数据来选择 K。
- 投票：K-NN 使用 K 个最近邻居中的多数投票来确定对象的类别。例如，如果 K 个邻居中的大多数属于 A 类，则该对象也将被归类为 A 类。
K-NN Classifier 的优点：
- 简单直观：K-NN是一种简单、直观的方法，易于理解和应用。
- 能够处理不同类型的数据类型：K-NN 可用于各种数据类型，包括数字、分类和文本数据。
- 对变化数据的适应性：K-NN 可以快速适应数据的变化，适合动态数据的任务。
K-NN Classifier 的局限性：
- 对 K 选择的敏感性：选择最佳 K 值并非易事。较小的 K 可能导致过度拟合，而较大的 K 可能导致欠拟合。
- 对特征缩放的敏感度：K-NN 对特征缩放很敏感，因此数据规范化非常重要。
- 计算复杂度：对于大型数据集和大量特征，计算所有对象对之间的距离可能会非常耗费计算资源。
- 缺乏可解释性：K-NN 结果的解释比较困难，尤其是当 K 很大且数据量很大时。
K-NN Classifier 是一种机器学习方法，可用于对象接近度至关重要的任务，例如推荐系统、文本分类和模式识别。它非常适合初始数据分析和快速模型原型设计。
2.11.1.K-Nearest Neighbors (K-NN) Classifier 模型创建代码
此代码演示了在 Iris 数据集上训练 K-Nearest Neighbors (K-NN) Classifier 模型、以 ONNX 格式导出以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_KNearestNeighborsClassifier.py # The code uses the K-Nearest Neighbors (KNN) Classifier for the Iris dataset, converts the model to ONNX format, saves it, and evaluates its accuracy. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.neighbors import KNeighborsClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a K-Nearest Neighbors (KNN) Classifier model knn_model = KNeighborsClassifier(n_neighbors=3) # train the model on the entire dataset knn_model.fit(X, y) # predict classes for the entire dataset y_pred = knn_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of KNN Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(knn_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "knn_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of KNN Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of KNN Classifier model:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.94 0.94 50
Python 2 0.94 0.94 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\knn_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of KNN Classifier model in ONNX format:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.94 0.94 50
Python 2 0.94 0.94 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\knn_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of KNN Classifier model in ONNX format:0.96
2.11.2.用于处理 K-Nearest Neighbors (K-NN) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_KNearestNeighborsClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "knn_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="KNearestNeighborsClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier correct results: 96.00% Iris_KNearestNeighborsClassifier (EURUSD,H1) model=KNearestNeighborsClassifier all samples accuracy=0.960000 Iris_KNearestNeighborsClassifier (EURUSD,H1) model:KNearestNeighborsClassifier FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50) Iris_KNearestNeighborsClassifier (EURUSD,H1) model=KNearestNeighborsClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为96%，与原始模型的准确率一致。
2.11.3.K-Nearest Neighbors (K-NN) Classifier 的 ONNX 表示
图 25.Netron 中 K-Nearest Neighbors 的 ONNX 表示
2.12.Decision Tree Classifier
Decision Tree Classifier（决策树分类器）是一种基于决策树构建的分类任务的机器学习方法。该方法通过对特征执行一系列条件测试将数据集划分为更小的子组，并根据对象在树中遵循的路径确定对象的类别。
Decision Tree Classifier 的原理：
- 构建决策树：最初，所有数据都表示在树的根部。对于树的每个节点，数据根据一个特征的值被分成两个或多个子组，旨在最小化每个子组中的不确定性（例如，熵或基尼指数）。
- 递归构造：分割数据的过程以递归方式进行，直到树到达其叶子为止。叶子代表对象的最终类别。
- 决策：当一个对象进入树时，它会沿着从根到其中一个叶子的路径移动，并根据该叶子中的大多数对象确定其类别。
Decision Tree Classifier 的优点：
- 可解释性：决策树易于解释和可视化。用于分类的决策规则是可以理解的。
- 处理不同的数据类型：Decision Tree Classifier 可以处理数字特征和分类特征。
- 特征重要性：决策树可以评估特征重要性，帮助数据分析师和特征工程师理解数据。
Decision Tree Classifier 的局限性：
- 过度拟合：大而深的树容易过度拟合，使得它们不太适合推广到新数据。
- 对噪声的敏感度：决策树对数据中的噪声和异常值很敏感。
- 贪婪构造：决策树是使用贪婪算法构建的，这可能导致次优的全局解。
- 数据变化的不稳定性：数据的微小变化可能会导致树结构的重大变化。
Decision Tree Classifier 是一种用于分类任务的有用的机器学习方法，特别是在模型可解释性至关重要且您需要了解哪些特征会影响决策的情况下。该方法还可以用于随机森林和梯度提升等集成方法。
2.12.1.创建 Decision Tree Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Decision Tree Classifier 模型、以 ONNX 格式导出以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_DecisionTreeClassifier.py # The code uses the Decision Tree Classifier for the Iris dataset, converts the model to ONNX format, saves it, and evaluates its accuracy. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.tree import DecisionTreeClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Decision Tree Classifier model decision_tree_model = DecisionTreeClassifier(random_state=42) # train the model on the entire dataset decision_tree_model.fit(X, y) # predict classes for the entire dataset y_pred = decision_tree_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Decision Tree Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(decision_tree_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "decision_tree_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Decision Tree Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Decision Tree Classifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\decision_tree_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Decision Tree Classifier model in ONNX format:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\decision_tree_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Decision Tree Classifier model in ONNX format:1.0
Decision Tree Classifier 模型（及其 ONNX 版本）在对完整 Fisher 鸢尾花数据集进行分类时表现出 100% 的准确率。
2.12.2.用于处理 Decision Tree Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_DecisionTreeClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "decision_tree_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="DecisionTreeClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_DecisionTreeClassifier (EURUSD,H1) model:DecisionTreeClassifier correct results: 100.00% Iris_DecisionTreeClassifier (EURUSD,H1) model=DecisionTreeClassifier all samples accuracy=1.000000 Iris_DecisionTreeClassifier (EURUSD,H1) model=DecisionTreeClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率一致。
2.12.3.Decision Tree Classifier 的 ONNX 表示
图 26.Netron 中 Decision Tree Classifier 的 ONNX 表示
关于 LogisticRegression 和 LogisticRegressionCV 的注意事项：
LogisticRegression 和 LogisticRegressionCV 是两个使用逻辑回归进行二元分类的分类器，但它们在模型参数的调整方式上有所不同：
LogisticRegression：
- LogisticRegression 是一种分类器，它使用逻辑函数来模拟属于两个类别之一的概率（二元分类）。
- 它提供自定义的基本参数，例如 C（逆正则化强度）、penalty（正则化类型，例如 L1 或 L2）、solver（优化算法）等。
- 使用 LogisticRegression 时，您通常会手动选择参数值及其组合，然后在数据上训练模型。
LogisticRegressionCV：
- LogisticRegressionCV 是 LogisticRegression 的扩展，它提供对交叉验证和选择正则化参数 C 的最优值的内置支持。
- 您不需要手动选择 C，而是可以向 LogisticRegressionCV 传递一个 C 值列表来探索并指定交叉验证方法（例如，通过 cv 参数）。
- LogisticRegressionCV 自动选择在交叉验证中表现最佳的最优 C 值。
- 当您需要自动调整正则化时，这非常方便，特别是当您有大量数据或不确定选择哪个 C 值时。
所以，它们之间的主要区别在于参数调整的自动化程度。LogisticRegression 需要手动调整C，而LogisticRegressionCV 可以通过交叉验证自动选择最佳的C值。它们之间的选择取决于您的需求以及对模型调整过程自动化的渴望。
2.13.Logistic Regression Classifier
Logistic Regression Classifier（逻辑回归分类器）是一种用于二元和多类分类任务的机器学习方法。尽管逻辑回归名为“回归”，但它实际上预测的是某个对象属于某一类别的概率。根据这些概率做出最终的分类决定。
Logistic Regression Classifier 的原理：
- 概率预测：逻辑回归使用逻辑（S 型）函数对对象属于特定类别的概率进行建模。
- 决策边界：根据预测的概率，逻辑回归确定区分类别的决策边界。如果概率超过某个阈值（通常为 0.5），则该对象被归类为一类；否则，则被归类为另一类。
- 参数学习：通过调整与特征相关的权重（系数）来最小化损失函数，在训练数据集上训练逻辑回归模型。
Logistic Regression Classifier 的优点：
- 简单性和可解释性：逻辑回归是一种简单的模型，其关于特征对类别预测的影响的结果很容易解释。
- 大型数据集的效率：逻辑回归可以有效地处理大型数据集并对其进行快速训练。
- 集成方法中的用法：逻辑回归可以作为堆叠等集成方法中的基础分类器。
Logistic Regression Classifier 的局限性：
- 线性：逻辑回归假设特征和几率的对数之间存在线性关系，这可能不适合复杂的任务。
- 多类约束：逻辑回归的原始形式是为二元分类而设计的，但有一些方法（如 One-vs-All（One-vs-Rest））可以将其扩展为多类分类。
- 对异常值的敏感性：逻辑回归对数据中的异常值很敏感。
逻辑回归是一种经典的机器学习方法，在实践中广泛用于分类任务，尤其是当模型的可解释性很重要，并且数据呈现线性或近线性结构时。它还用于统计和医学数据分析，以评估因素对事件发生可能性的影响。
2.13.1.创建 Logistic Regression Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Logistic Regression Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_LogisticRegressionClassifier.py # The code uses the Logistic Regression Classifier for the Iris dataset, converts the model to ONNX format, saves it, and evaluates its accuracy. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import LogisticRegression from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Logistic Regression Classifier model logistic_regression_model = LogisticRegression(max_iter=1000, random_state=42) # train the model on the entire dataset logistic_regression_model.fit(X, y) # predict classes for the entire dataset y_pred = logistic_regression_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Logistic Regression Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(logistic_regression_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "logistic_regression_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Logistic Regression Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Logistic Regression Classifier model:0.9733333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.94 0.96 50
Python 2 0.94 0.98 0.96 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\logistic_regression_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Logistic Regression Classifier model in ONNX format:0.9733333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.94 0.96 50
Python 2 0.94 0.98 0.96 50
Python
Python accuracy 0.97 150
Python macro avg 0.97 0.97 0.97 150
Python weighted avg 0.97 0.97 0.97 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\logistic_regression_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Logistic Regression Classifier model in ONNX format:0.9733333333333334
2.13.2.用于处理 Logistic Regression Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_LogisticRegressionClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "logistic_regression_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="LogisticRegressionClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+输出：
Iris_LogisticRegressionClassifier (EURUSD,H1) model:LogisticRegressionClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_LogisticRegressionClassifier (EURUSD,H1) model:LogisticRegressionClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_LogisticRegressionClassifier (EURUSD,H1) model:LogisticRegressionClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_LogisticRegressionClassifier (EURUSD,H1) model:LogisticRegressionClassifier sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_LogisticRegressionClassifier (EURUSD,H1) model:LogisticRegressionClassifier correct results: 97.33% Iris_LogisticRegressionClassifier (EURUSD,H1) model=LogisticRegressionClassifier all samples accuracy=0.973333 Iris_LogisticRegressionClassifier (EURUSD,H1) model=LogisticRegressionClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 97.33%，与原始模型的准确率一致。
2.13.3.Logistic Regression Classifier 的 ONNX 表示
图 27.Netron 中 Logistic Regression Classifier 的 ONNX 表示
2.14.LogisticRegressionCV Classifier
LogisticRegressionCV（带交叉验证的逻辑回归）是一种强大且灵活的二元分类方法。该方法不仅允许您创建基于逻辑回归的分类模型，还可以自动调整参数以获得最佳性能。
LogisticRegressionCV 的工作原理：
- 逻辑回归：LogisticRegressionCV 从根本上来说基于逻辑回归。逻辑回归是一种统计方法，用于模拟一个对象属于两个类别之一的概率。当因变量为二进制（两个类）或可以转换为二进制时，适用此模型。
- 交叉验证：LogisticRegressionCV 的关键优势在于其集成的交叉验证。这意味着该方法不需要手动选择正则化参数 C 的最优值，而是自动尝试不同的 C 值并选择在交叉验证中表现最佳的 C 值。
- 选择最佳C：LogisticRegressionCV 采用交叉验证策略来评估模型在不同 C 值下的性能。C 是控制模型正则化程度的正则化参数。C 值较小表示正则化程度强，C 值较大表示正则化程度弱。交叉验证有助于选择最佳 C 值来平衡欠拟合和过度拟合。
- 正则化：LogisticRegressionCV 还支持各种类型的正则化，包括 L1（lasso，套索）和 L2（ridge，岭）正则化。这些正则化类型有助于提高模型的泛化能力并防止过度拟合。
LogisticRegressionCV 的优点：
自动参数调整：LogisticRegressionCV 的主要优势之一是它能够使用交叉验证自动选择最佳 C 值。这消除了手动模型调整的需要，使您能够专注于数据和任务。
过度拟合稳健性：LogisticRegressionCV 支持的正则化有助于控制模型的复杂性并降低过度拟合的风险，尤其是在数据有限的情况下。
透明度：逻辑回归是一种可解释的方法。您可以分析每个特征对预测的贡献，这对于理解特征重要性很有用。
高性能：逻辑回归可以快速有效地工作，特别是在处理大量数据时。
LogisticRegressionCV 的局限性：
线性相关性：LogisticRegressionCV 适合解决线性和近线性分类问题。如果特征和目标变量之间的关系高度非线性，则模型可能无法表现良好。
处理大量特征：由于特征数量众多，逻辑回归可能需要大量数据或降维技术来防止过度拟合。
数据表示依赖性：逻辑回归的有效性取决于数据的表示方式和选择的特征。
LogisticRegressionCV 是一个强大的二元分类工具，具有自动参数调整和过度拟合鲁棒性。当您需要快速构建可解释的分类模型时它特别有用。然而，重要的是要记住，当数据表现出线性或近线性依赖性时，它表现最佳。
2.14.1.创建 LogisticRegressionCV Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 LogisticRegressionCV Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_LogisticRegressionCVClassifier.py # The code demonstrates the process of training LogisticRegressionCV model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import LogisticRegressionCV from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a LogisticRegressionCV model logistic_regression_model = LogisticRegressionCV(cv=5, max_iter=1000) # train the model on the entire dataset logistic_regression_model.fit(X, y) # predict classes for the entire dataset y_pred = logistic_regression_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of LogisticRegressionCV model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(logistic_regression_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "logistic_regressioncv_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of LogisticRegressionCV model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of LogisticRegressionCV model:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.96 0.97 50
Python 2 0.96 0.98 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\logistic_regression_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of LogisticRegressionCV model in ONNX format:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.96 0.97 50
Python 2 0.96 0.98 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\logistic_regression_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of LogisticRegressionCV model in ONNX format:0.98
2.14.2.用于处理 LogisticRegressionCV Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_LogisticRegressionCVClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "logistic_regressioncv_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="LogisticRegressionCVClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_LogisticRegressionCVClassifier (EURUSD,H1) model:LogisticRegressionCVClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_LogisticRegressionCVClassifier (EURUSD,H1) model:LogisticRegressionCVClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_LogisticRegressionCVClassifier (EURUSD,H1) model:LogisticRegressionCVClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_LogisticRegressionCVClassifier (EURUSD,H1) model:LogisticRegressionCVClassifier correct results: 98.00% Iris_LogisticRegressionCVClassifier (EURUSD,H1) model=LogisticRegressionCVClassifier all samples accuracy=0.980000 Iris_LogisticRegressionCVClassifier (EURUSD,H1) model=LogisticRegressionCVClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 98%，与原始模型的准确率一致。
2.14.3.LogisticRegressionCV Classifier 的 ONNX 表示
图 28.Netron 中 LogisticRegressionCV Classifier 的 ONNX 表示
2.15.Passive-Aggressive (PA) Classifier
Passive-Aggressive (PA) Classifier （被动攻击分类器）是一种用于分类任务的机器学习方法。该方法的核心思想是在训练过程中调整模型的权重（系数），以最大限度地减少分类误差。Passive-Aggressive Classifier 在在线学习场景和数据随时间变化的情况下很有用。
Passive-Aggressive Classifier 的工作原理：
- 权重调整：Passive-Aggressive Classifier 不会像随机梯度下降那样以最小化损失函数的方向更新模型的权重，而是以最小化当前示例的分类误差的方向调整权重。
- 保持攻击性：该方法有一个称为攻击性（C）的参数，它决定了模型权重的调整强度。C 值越大，该方法的适应性越强，而 C 值越小，该方法的适应性越弱。
Passive-Aggressive Classifier 的优点：
- 适合在线学习：Passive-Aggressive Classifier 可以随着新数据的到达而更新，使其适合数据以流形式到达的在线学习任务。
- 对数据变化的适应性：由于该方法能够使模型适应新的情况，因此在数据变化时表现良好。
Passive-Aggressive Classifier 的局限性：
- 攻击性敏感性参数选择：选择攻击性参数 C 的最佳值可能需要调整，并且取决于数据特征。
- 并不总是适合复杂的任务：在需要考虑复杂特征依赖关系的复杂任务中，Passive-Aggressive Classifier 可能无法提供高精度。
- 权重的解释：与使用线性或逻辑回归获得的权重相比，使用此方法获得的模型权重可能更难解释。
Passive-Aggressive Classifier 是一种机器学习方法，适用于数据不断变化的分类任务，以及模型快速适应新情况至关重要的情况。它可以应用于各个领域，包括文本数据分析、图像分类和其他任务。
2.15.1.创建 Passive-Aggressive (PA) Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Passive-Aggressive (PA) Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_PassiveAgressiveClassifier.py # The code uses the Passive-Aggressive (PA) Classifier for the Iris dataset, converts the model to ONNX format, saves it, and evaluates its accuracy. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import PassiveAggressiveClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Passive-Aggressive (PA) Classifier model pa_classifier_model = PassiveAggressiveClassifier(max_iter=1000, random_state=42) # train the model on the entire dataset pa_classifier_model.fit(X, y) # predict classes for the entire dataset y_pred = pa_classifier_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Passive-Aggressive (PA) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(pa_classifier_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "pa_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Passive-Aggressive (PA) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Passive-Aggressive (PA) Classifier model:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.96 0.92 0.94 50
Python 2 0.92 0.96 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\pa_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Passive-Aggressive (PA) Classifier model in ONNX format:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.96 0.92 0.94 50
Python 2 0.92 0.96 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\pa_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Passive-Aggressive (PA) Classifier model in ONNX format:0.96
2.15.2.用于处理 Passive-Aggressive (PA) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_PassiveAgressiveClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "pa_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="PassiveAgressiveClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_PassiveAgressiveClassifier (EURUSD,H1) model:PassiveAgressiveClassifier correct results: 96.00% Iris_PassiveAgressiveClassifier (EURUSD,H1) model=PassiveAgressiveClassifier all samples accuracy=0.960000 Iris_PassiveAgressiveClassifier (EURUSD,H1) model=PassiveAgressiveClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 96%，与原始模型的准确率一致。
2.15.3.Passive-Aggressive (PA) Classifier 的 ONNX 表示
图 29.Netron 中 Passive-Aggressive (PA) Classifier 的 ONNX 表示
2.16.Perceptron Classifier
Perceptron Classifier（感知器分类器）是一个线性二元分类器，用于根据线性分离超平面分离两个类。它是最简单、最古老的机器学习方法之一，其核心原理是训练模型的权重（系数），以最大化训练数据集上的分类准确率。
Perceptron Classifier 的工作原理：
- 线性超平面：感知器在特征空间中构建一个线性超平面，将两个类分开。该超平面由模型的权重（系数）决定。
- 权重训练：最初，权重被随机初始化或初始化为零。然后，对于训练数据集中的每个对象，模型根据当前权重预测类别，并在出现错误时进行调整。训练持续进行，直到所有对象都被正确分类或达到最大迭代次数。
Perceptron Classifier 的优点：
- 简单性：感知器是一种非常简单的算法，易于理解和实现。
- 训练速度快：感知器可以快速训练，尤其是在大型数据集上，并且可以用于在线学习任务。
Perceptron Classifier 的局限性：
- 线性可分离性约束：感知器仅在数据线性可分的情况下才能发挥良好的作用。如果数据不能线性分离，感知器可能无法达到高精度。
- 对初始权重的敏感度：权重的初始选择会影响算法的收敛。初始权重选择不当可能会导致收敛速度慢或神经元无法正确分类。
- 无法确定概率：感知器不提供类别成员的概率估计，这对于某些任务来说可能很重要。
感知器分类器是二元分类的基本算法，在数据线性可分的简单情况下很有用。它还可以作为更复杂方法的基础，例如多层神经网络。重要的是要记住，在数据具有复杂结构的更复杂的任务中，逻辑回归或支持向量机（SVM）等其他方法可以提供更高的分类准确性。
2.16.1.创建 Perceptron Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Perceptron Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_PerceptronClassifier.py # The code demonstrates the process of training Perceptron Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import Perceptron from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Perceptron Classifier model perceptron_model = Perceptron(max_iter=1000, random_state=42) # train the model on the entire dataset perceptron_model.fit(X, y) # predict classes for the entire dataset y_pred = perceptron_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Perceptron Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(perceptron_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "perceptron_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Perceptron Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Perceptron Classifier model:0.6133333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 0.80 0.89 50
Python 1 0.46 1.00 0.63 50
Python 2 1.00 0.04 0.08 50
Python
Python accuracy 0.61 150
Python macro avg 0.82 0.61 0.53 150
Python weighted avg 0.82 0.61 0.53 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\perceptron_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Perceptron Classifier model in ONNX format:0.6133333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 0.80 0.89 50
Python 1 0.46 1.00 0.63 50
Python 2 1.00 0.04 0.08 50
Python
Python accuracy 0.61 150
Python macro avg 0.82 0.61 0.53 150
Python weighted avg 0.82 0.61 0.53 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\perceptron_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Perceptron Classifier model in ONNX format:0.6133333333333333
2.16.2.用于处理 Perceptron Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_PerceptronClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "perceptron_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="PerceptronClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=2 FAILED [class=1, true class=0] features=(4.90,3.00,1.40,0.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=9 FAILED [class=1, true class=0] features=(4.40,2.90,1.40,0.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=10 FAILED [class=1, true class=0] features=(4.90,3.10,1.50,0.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=13 FAILED [class=1, true class=0] features=(4.80,3.00,1.40,0.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=21 FAILED [class=1, true class=0] features=(5.40,3.40,1.70,0.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=26 FAILED [class=1, true class=0] features=(5.00,3.00,1.60,0.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=31 FAILED [class=1, true class=0] features=(4.80,3.10,1.60,0.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=35 FAILED [class=1, true class=0] features=(4.90,3.10,1.50,0.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=42 FAILED [class=1, true class=0] features=(4.50,2.30,1.30,0.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=46 FAILED [class=1, true class=0] features=(4.80,3.00,1.40,0.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=102 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=103 FAILED [class=1, true class=2] features=(7.10,3.00,5.90,2.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=104 FAILED [class=1, true class=2] features=(6.30,2.90,5.60,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=105 FAILED [class=1, true class=2] features=(6.50,3.00,5.80,2.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=106 FAILED [class=1, true class=2] features=(7.60,3.00,6.60,2.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=108 FAILED [class=1, true class=2] features=(7.30,2.90,6.30,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=109 FAILED [class=1, true class=2] features=(6.70,2.50,5.80,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=110 FAILED [class=1, true class=2] features=(7.20,3.60,6.10,2.50] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=111 FAILED [class=1, true class=2] features=(6.50,3.20,5.10,2.00] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=112 FAILED [class=1, true class=2] features=(6.40,2.70,5.30,1.90] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=113 FAILED [class=1, true class=2] features=(6.80,3.00,5.50,2.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=114 FAILED [class=1, true class=2] features=(5.70,2.50,5.00,2.00] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=116 FAILED [class=1, true class=2] features=(6.40,3.20,5.30,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=117 FAILED [class=1, true class=2] features=(6.50,3.00,5.50,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=118 FAILED [class=1, true class=2] features=(7.70,3.80,6.70,2.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=119 FAILED [class=1, true class=2] features=(7.70,2.60,6.90,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=121 FAILED [class=1, true class=2] features=(6.90,3.20,5.70,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=122 FAILED [class=1, true class=2] features=(5.60,2.80,4.90,2.00] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=123 FAILED [class=1, true class=2] features=(7.70,2.80,6.70,2.00] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=124 FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=125 FAILED [class=1, true class=2] features=(6.70,3.30,5.70,2.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=126 FAILED [class=1, true class=2] features=(7.20,3.20,6.00,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=128 FAILED [class=1, true class=2] features=(6.10,3.00,4.90,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=129 FAILED [class=1, true class=2] features=(6.40,2.80,5.60,2.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=131 FAILED [class=1, true class=2] features=(7.40,2.80,6.10,1.90] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=132 FAILED [class=1, true class=2] features=(7.90,3.80,6.40,2.00] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=133 FAILED [class=1, true class=2] features=(6.40,2.80,5.60,2.20] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=136 FAILED [class=1, true class=2] features=(7.70,3.00,6.10,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=137 FAILED [class=1, true class=2] features=(6.30,3.40,5.60,2.40] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=138 FAILED [class=1, true class=2] features=(6.40,3.10,5.50,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=140 FAILED [class=1, true class=2] features=(6.90,3.10,5.40,2.10] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=141 FAILED [class=1, true class=2] features=(6.70,3.10,5.60,2.40] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=142 FAILED [class=1, true class=2] features=(6.90,3.10,5.10,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=143 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=144 FAILED [class=1, true class=2] features=(6.80,3.20,5.90,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=145 FAILED [class=1, true class=2] features=(6.70,3.30,5.70,2.50] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=146 FAILED [class=1, true class=2] features=(6.70,3.00,5.20,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=147 FAILED [class=1, true class=2] features=(6.30,2.50,5.00,1.90] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=148 FAILED [class=1, true class=2] features=(6.50,3.00,5.20,2.00] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=149 FAILED [class=1, true class=2] features=(6.20,3.40,5.40,2.30] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier sample=150 FAILED [class=1, true class=2] features=(5.90,3.00,5.10,1.80] Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier correct results: 61.33% Iris_PerceptronClassifier (EURUSD,H1) model=PerceptronClassifier all samples accuracy=0.613333 Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80) Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier FAILED [class=1, true class=0] features=(4.90,3.10,1.50,0.10) Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90) Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier FAILED [class=1, true class=2] features=(7.10,3.00,5.90,2.10) Iris_PerceptronClassifier (EURUSD,H1) model:PerceptronClassifier FAILED [class=1, true class=2] features=(6.30,2.90,5.60,1.80) Iris_PerceptronClassifier (EURUSD,H1) model=PerceptronClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 61.33%，与原始模型的准确率相对应。
2.16.3.Perceptron Classifier 的 ONNX 表示
图 30.Netron 中 Perceptron Classifier 的 ONNX 表示
2.17.Stochastic Gradient Descent Classifier
SGD Classifier（随机梯度下降分类器，Stochastic Gradient Descent Classifier）是一种用于分类任务的机器学习方法。它是线性模型的一个特例，是使用随机梯度下降训练的线性分类器。
SGD Classifier 的原理：
- 线性超平面：SGD 分类器在多维特征空间中构建一个线性超平面，将两个类分开。该超平面由模型的权重（系数）决定。
- 随机梯度下降：该方法使用随机梯度下降进行训练，这意味着对训练数据集（或随机选择的子集）中的每个对象而不是整个数据集执行权重更新。这使得 SGD Classifier 适合大量数据和在线学习。
- 损失函数：SGD Classifier 优化损失函数，例如用于二元分类的逻辑损失函数或用于多类分类的 softmax 损失函数。
SGD Classifier 的优点：
- 训练速度：由于随机梯度下降，SGD 分类器训练速度很快，特别是在大量数据的情况下。
- 适合在线学习：该方法非常适合在线学习任务，其中数据以流式方式到达，并且模型需要随着新数据的进入而更新。
SGD Classifier 的局限性：
- 对参数的敏感性：SGD Classifier 有许多超参数，例如学习率和正则化参数，需要仔细调整。
- 权重初始化：权重的初始选择会影响收敛和模型质量。
- 收敛到局部最小值：由于SGD方法的随机性，它会收敛到损失函数的局部最小值，这会影响模型质量。
SGD Classifier 是一种多功能的机器学习方法，可用于二元和多类分类任务，尤其是在处理需要快速处理的大量数据时。适当调整其超参数并监控收敛以实现高分类准确性非常重要。
2.17.1.创建 Stochastic Gradient Descent Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 SGD Classfier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_SGDClassifier.py # The code demonstrates the process of training Stochastic Gradient Descent Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.linear_model import SGDClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create an SGD Classifier model sgd_model = SGDClassifier(max_iter=1000, random_state=42) # train the model on the entire dataset sgd_model.fit(X, y) # predict classes for the entire dataset y_pred = sgd_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of SGD Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(sgd_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "sgd_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of SGD Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of SGD Classifier model:0.9333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.96 1.00 0.98 50
Python 1 0.88 0.92 0.90 50
Python 2 0.96 0.88 0.92 50
Python
Python accuracy 0.93 150
Python macro avg 0.93 0.93 0.93 150
Python weighted avg 0.93 0.93 0.93 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\perceptron_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of SGD Classifier model in ONNX format:0.9333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.96 1.00 0.98 50
Python 1 0.88 0.92 0.90 50
Python 2 0.96 0.88 0.92 50
Python
Python accuracy 0.93 150
Python macro avg 0.93 0.93 0.93 150
Python weighted avg 0.93 0.93 0.93 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\perceptron_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of SGD Classifier model in ONNX format:0.9333333333333333
2.17.2.用于处理 Stochastic Gradient Descent Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_SGDClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "sgd_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="SGDClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=65 FAILED [class=0, true class=1] features=(5.60,2.90,3.60,1.30] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=86 FAILED [class=0, true class=1] features=(6.00,3.40,4.50,1.60] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=124 FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier correct results: 93.33% Iris_SGDClassifier (EURUSD,H1) model=SGDClassifier all samples accuracy=0.933333 Iris_SGDClassifier (EURUSD,H1) model:SGDClassifier FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80) Iris_SGDClassifier (EURUSD,H1) model=SGDClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 93.33%，与原始模型的准确率一致。
2.17.3.Stochastic Gradient Descent Classifier 的 ONNX 表示
图 31.Netron 中 Stochastic Gradient Descent Classifier 的 ONNX 表示
2.18.Gaussian Naive Bayes (GNB) Classifier
Gaussian Naive Bayes (GNB，高斯朴素贝叶斯) Classifier 是一种基于贝叶斯概率模型的机器学习方法，用于分类任务。它是朴素贝叶斯分类器家族的一部分，假设所有特征都是独立的并且具有正态分布。
Gaussian Naive Bayes Classifier 的原理：
- 贝叶斯方法：GNB 基于贝叶斯分类方法，该方法使用贝叶斯定理来计算一个对象属于每个类别的概率。
- 朴素假设：GNB 中的关键假设是所有特征都是独立的并且遵循正态（高斯）分布。这种假设被认为是幼稚的，因为在现实世界的数据中，特征通常相互关联。
- 参数估计：GNB 模型在训练数据集上进行训练，通过计算每个类别中每个特征的分布参数（平均值和标准差）。
Gaussian Naive Bayes Classifier 的优点：
- 简单性和训练速度：GNB 是一种非常简单的算法，即使在大型数据集上也能快速训练。
- 对中小型数据的有效性：GNB 对于具有少量或中等数量特征的分类任务非常有效，尤其是当正态特征分布假设成立时。
Gaussian Naive Bayes Classifier 的局限性：
- 朴素假设：特征独立性和正态分布的假设对于现实世界的数据来说可能过于简单和不正确，从而导致分类准确性降低。
- 对异常值的敏感性：GNB 对数据中的异常值很敏感，因为它们会严重扭曲正态分布的参数。
- 无法捕获特征依赖关系：由于独立性假设，GNB 不考虑特征之间的依赖关系。
Gaussian Naive Bayes Classifier 对于简单的分类任务来说是一个不错的选择，特别是当近似满足正态特征分布的假设时。然而，在更复杂的任务中，特征是相关或不遵循正态分布，其他方法（如支持向量机（SVM）或梯度提升）可以提供更准确的结果。
2.18.1.创建 Gaussian Naive Bayes (GNB) Classifier 模型的代码
此代码演示了在 Iris 数据集上训练Gaussian Naive Bayes (GNB) Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_GaussianNaiveBayesClassifier.py # The code demonstrates the process of training Gaussian Naive Bayes Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.naive_bayes import GaussianNB from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Gaussian Naive Bayes (GNB) Classifier model gnb_model = GaussianNB() # train the model on the entire dataset gnb_model.fit(X, y) # predict classes for the entire dataset y_pred = gnb_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Gaussian Naive Bayes (GNB) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(gnb_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "gnb_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Gaussian Naive Bayes (GNB) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Gaussian Naive Bayes (GNB) Classifier model:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.94 0.94 50
Python 2 0.94 0.94 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\gnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Gaussian Naive Bayes (GNB) Classifier model in ONNX format:0.96
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.94 0.94 50
Python 2 0.94 0.94 0.94 50
Python
Python accuracy 0.96 150
Python macro avg 0.96 0.96 0.96 150
Python weighted avg 0.96 0.96 0.96 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\gnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Gaussian Naive Bayes (GNB) Classifier model in ONNX format:0.96
2.18.2.用于处理 Gaussian Naive Bayes (GNB) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_GaussianNaiveBayesClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "gnb_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="GaussianNaiveBayesClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model:GaussianNaiveBayesClassifier correct results: 96.00% Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model=GaussianNaiveBayesClassifier all samples accuracy=0.960000 Iris_GaussianNaiveBayesClassifier (EURUSD,H1) model=GaussianNaiveBayesClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 96%，与原始模型的准确率一致。
2.18.3.Gaussian Naive Bayes (GNB) Classifier 的 ONNX 表示
图 32.Netron 中 Gaussian Naive Bayes (GNB) Classifier 的 ONNX 表示
2.19.Multinomial Naive Bayes (MNB) Classifier
Multinomial Naive Bayes (MNB，多项式朴素贝叶斯 ) Classifier 是一种基于贝叶斯概率模型的机器学习方法，用于分类任务，尤其是在文本处理中。它是朴素贝叶斯分类器的变体之一，假设特征代表计数，例如文本中单词出现的次数。
Multinomial Naive Bayes Classifier 的原理：
- 贝叶斯方法：MNB 也遵循贝叶斯分类方法，使用贝叶斯定理计算一个对象属于每个类别的概率。
- 多项分布的假设：MNB 中的主要假设是特征代表计数，例如文本中单词出现的次数，并遵循多项分布。该假设通常对于文本数据有效。
- 参数估计：MNB 模型在训练数据集上通过计算每个类别中每个特征的分布参数进行训练。
Multinomial Naive Bayes Classifier 的优点：
- 文本处理的有效性：由于特征计数的假设，MNB 在与文本数据分析相关的任务（例如文本分类或垃圾邮件过滤）中表现良好。
- 简单性和训练速度：与其他朴素贝叶斯分类器一样，MNB 是一种简单的算法，即使在大量文本数据上也能快速训练。
Multinomial Naive Bayes Classifier 的局限性：
- 朴素假设：对于现实世界的数据来说，特征的多项分布假设可能过于简单且不准确，尤其是当特征具有复杂结构时。
- 无法解释词序：MNB 不考虑文本中单词的顺序，这在某些文本分析任务中很重要。
- 对生僻词的敏感度：MNB 对罕见词很敏感，出现次数不足会降低分类准确性。
Multinomial Naive Bayes Classifier 是文本分析任务的一种有用方法，特别是当特征与计数相关时，例如文本中的单词数。它广泛用于自然语言处理 (NLP) 的文本分类、文档分类和其他文本分析。
2.19.1.创建 Multinomial Naive Bayes (MNB) Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Multinomial Naive Bayes (MNB) Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_MultinomialNaiveBayesClassifier.py # The code demonstrates the process of training Multinomial Naive Bayes (MNB) Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.naive_bayes import MultinomialNB from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Multinomial Naive Bayes (MNB) Classifier model mnb_model = MultinomialNB() # train the model on the entire dataset mnb_model.fit(X, y) # predict classes for the entire dataset y_pred = mnb_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Multinomial Naive Bayes (MNB) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(mnb_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "mnb_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Multinomial Naive Bayes (MNB) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Multinomial Naive Bayes (MNB) Classifier model:0.9533333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.92 0.93 50
Python 2 0.92 0.94 0.93 50
Python
Python accuracy 0.95 150
Python macro avg 0.95 0.95 0.95 150
Python weighted avg 0.95 0.95 0.95 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\mnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Multinomial Naive Bayes (MNB) Classifier model in ONNX format:0.9533333333333334
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.94 0.92 0.93 50
Python 2 0.92 0.94 0.93 50
Python
Python accuracy 0.95 150
Python macro avg 0.95 0.95 0.95 150
Python weighted avg 0.95 0.95 0.95 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\mnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Multinomial Naive Bayes (MNB) Classifier model in ONNX format:0.9533333333333334
2.19.2.用于处理 Multinomial Naive Bayes (MNB) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_MultinomialNaiveBayesClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "mnb_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="MultinomialNaiveBayesClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=69 FAILED [class=2, true class=1] features=(6.20,2.20,4.50,1.50] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=132 FAILED [class=1, true class=2] features=(7.90,3.80,6.40,2.00] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier correct results: 95.33% Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model=MultinomialNaiveBayesClassifier all samples accuracy=0.953333 Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model:MultinomialNaiveBayesClassifier FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50) Iris_MultinomialNaiveBayesClassifier (EURUSD,H1) model=MultinomialNaiveBayesClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 95.33%，与原始模型的准确率一致。
2.19.3.Multinomial Naive Bayes (MNB) Classifier 的 ONNX 表示
图 33.Netron 中的 Multinomial Naive Bayes (MNB) Classifier 的 ONNX 表示
2.20.Complement Naive Bayes (CNB) Classifier
Complement Naive Bayes (CNB，补充朴素贝叶斯) Classifier 是朴素贝叶斯分类器的一种变体，专门设计用于处理不平衡数据，其中一个类别可能比另一个类别更为普遍。该分类器采用经典的朴素贝叶斯方法来解决类别不平衡问题。
Complement Naive Bayes Classifier 的原理：
- 贝叶斯方法：与其他贝叶斯分类器一样，CNB 遵循贝叶斯分类方法，并使用贝叶斯定理计算对象属于每个类的概率。
- 解决类别不平衡（Class Imbalance）问题：CNB 的主要目的是纠正类别不平衡。与标准朴素贝叶斯方法不同，CNB 不会考虑类内特征的概率，而是尝试考虑类外特征的概率。当一个类别的表现性明显低于另一个类别时，这尤其有用。
- 参数估计：CNB 模型在训练数据集上通过计算类外每个特征的分布参数进行训练。
Complement Naive Bayes Classifier 的优点：
- 不平衡数据的适用性：CNB 在数据不平衡（类别具有不同的频率）的分类任务中表现良好。
- 简单性和训练速度：与其他朴素贝叶斯分类器一样，CNB 是一种简单的算法，即使在大量数据的情况下也可以快速训练。
Complement Naive Bayes Classifier 分类器的局限性：
- 对正则化参数选择的敏感性：与其他贝叶斯方法一样，选择正则化参数的正确值可能需要调整和评估。
- 朴素假设：与其他朴素贝叶斯分类器一样，CNB 做出了特征独立性的假设，这对于某些任务来说可能过于简单。
对于数据不平衡的分类任务， Complement Naive Bayes Classifier 是一个不错的选择，尤其是当一个类别的表现性明显低于另一个类别时。它在文本分类任务中特别有用，因为不同类别的单词可能严重不平衡，例如情感分析或垃圾邮件过滤。
2.20.1.创建 Complement Naive Bayes (CNB) Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Complement Naive Bayes (CNB) Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_CNBClassifier.py # The code demonstrates the process of training Complement Naive Bayes (CNB) Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.naive_bayes import ComplementNB from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Complement Naive Bayes (CNB) Classifier model cnb_model = ComplementNB() # train the model on the entire dataset cnb_model.fit(X, y) # predict classes for the entire dataset y_pred = cnb_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Complement Naive Bayes (CNB) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(cnb_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "cnb_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Complement Naive Bayes (CNB) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Complement Naive Bayes (CNB) Classifier model:0.6666666666666666
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.96 1.00 0.98 50
Python 1 0.00 0.00 0.00 50
Python 2 0.51 1.00 0.68 50
Python
Python accuracy 0.67 150
Python macro avg 0.49 0.67 0.55 150
Python weighted avg 0.49 0.67 0.55 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\cnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Complement Naive Bayes (CNB) Classifier model in ONNX format:0.6666666666666666
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.96 1.00 0.98 50
Python 1 0.00 0.00 0.00 50
Python 2 0.51 1.00 0.68 50
Python
Python accuracy 0.67 150
Python macro avg 0.49 0.67 0.55 150
Python weighted avg 0.49 0.67 0.55 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\cnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Complement Naive Bayes (CNB) Classifier model in ONNX format:0.6666666666666666
2.20.2.用于处理 Complement Naive Bayes (CNB) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_CNBClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "cnb_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="CNBClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=51 FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=52 FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=54 FAILED [class=2, true class=1] features=(5.50,2.30,4.00,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=55 FAILED [class=2, true class=1] features=(6.50,2.80,4.60,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=56 FAILED [class=2, true class=1] features=(5.70,2.80,4.50,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=57 FAILED [class=2, true class=1] features=(6.30,3.30,4.70,1.60] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=58 FAILED [class=2, true class=1] features=(4.90,2.40,3.30,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=59 FAILED [class=2, true class=1] features=(6.60,2.90,4.60,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=60 FAILED [class=2, true class=1] features=(5.20,2.70,3.90,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=61 FAILED [class=2, true class=1] features=(5.00,2.00,3.50,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=62 FAILED [class=2, true class=1] features=(5.90,3.00,4.20,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=63 FAILED [class=2, true class=1] features=(6.00,2.20,4.00,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=64 FAILED [class=2, true class=1] features=(6.10,2.90,4.70,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=65 FAILED [class=2, true class=1] features=(5.60,2.90,3.60,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=66 FAILED [class=2, true class=1] features=(6.70,3.10,4.40,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=68 FAILED [class=2, true class=1] features=(5.80,2.70,4.10,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=69 FAILED [class=2, true class=1] features=(6.20,2.20,4.50,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=70 FAILED [class=2, true class=1] features=(5.60,2.50,3.90,1.10] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=72 FAILED [class=2, true class=1] features=(6.10,2.80,4.00,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=74 FAILED [class=2, true class=1] features=(6.10,2.80,4.70,1.20] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=75 FAILED [class=2, true class=1] features=(6.40,2.90,4.30,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=76 FAILED [class=2, true class=1] features=(6.60,3.00,4.40,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=77 FAILED [class=2, true class=1] features=(6.80,2.80,4.80,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=79 FAILED [class=2, true class=1] features=(6.00,2.90,4.50,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=80 FAILED [class=0, true class=1] features=(5.70,2.60,3.50,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=81 FAILED [class=2, true class=1] features=(5.50,2.40,3.80,1.10] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=82 FAILED [class=2, true class=1] features=(5.50,2.40,3.70,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=83 FAILED [class=2, true class=1] features=(5.80,2.70,3.90,1.20] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=86 FAILED [class=2, true class=1] features=(6.00,3.40,4.50,1.60] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=87 FAILED [class=2, true class=1] features=(6.70,3.10,4.70,1.50] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=88 FAILED [class=2, true class=1] features=(6.30,2.30,4.40,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=89 FAILED [class=2, true class=1] features=(5.60,3.00,4.10,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=90 FAILED [class=2, true class=1] features=(5.50,2.50,4.00,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=91 FAILED [class=2, true class=1] features=(5.50,2.60,4.40,1.20] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=92 FAILED [class=2, true class=1] features=(6.10,3.00,4.60,1.40] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=93 FAILED [class=2, true class=1] features=(5.80,2.60,4.00,1.20] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=94 FAILED [class=2, true class=1] features=(5.00,2.30,3.30,1.00] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=95 FAILED [class=2, true class=1] features=(5.60,2.70,4.20,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=96 FAILED [class=2, true class=1] features=(5.70,3.00,4.20,1.20] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=97 FAILED [class=2, true class=1] features=(5.70,2.90,4.20,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=98 FAILED [class=2, true class=1] features=(6.20,2.90,4.30,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=99 FAILED [class=0, true class=1] features=(5.10,2.50,3.00,1.10] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier sample=100 FAILED [class=2, true class=1] features=(5.70,2.80,4.10,1.30] Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier correct results: 66.67% Iris_CNBClassifier (EURUSD,H1) model=CNBClassifier all samples accuracy=0.666667 Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50) Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40) Iris_CNBClassifier (EURUSD,H1) model:CNBClassifier FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50) Iris_CNBClassifier (EURUSD,H1) model=CNBClassifier batch test accuracy=0.000000导出的 ONNX 模型在完整 Iris 数据集上的准确率为 66.67%，与原始模型的准确率相对应。
2.20.3.Complement Naive Bayes (CNB) Classifier 的 ONNX 表示
图 34.Netron 中 Complement Naive Bayes (CNB) Classifier 的 ONNX 表示
2.21.Bernoulli Naive Bayes (BNB) Classifier
Bernoulli Naive Bayes (BNB，伯努利朴素贝叶斯) Classifier 是朴素贝叶斯分类器的另一种变体，用于二元分类任务。这种分类器在特征以二进制数据表示的情况下特别有用，例如在文本分析中，特征可能是文本中单词的存在或不存在。
Bernoulli Naive Bayes Classifier 的原理：
- 贝叶斯方法：与其他贝叶斯分类器一样，BNB 遵循贝叶斯分类方法，并使用贝叶斯定理来计算对象属于每个类别的概率。
- 二元特征假设：BNB 的主要假设是特征表示为二进制数据，这意味着它们只能有两个值，例如 1 和 0，其中 1 表示特征存在，0 表示特征不存在。
- 参数估计：BNB 模型在训练数据集上通过计算每个类别中每个特征的分布参数进行训练。
Bernoulli Naive Bayes Classifier 的优点：
- 二进制数据的有效性：BNB 在以二进制数据表示特征的任务中效果很好，并且在文本分析或事件分类中特别有用。
- 简单性和训练速度：与其他朴素贝叶斯分类器一样，BNB 是一种简单且训练快速的算法。
Bernoulli Naive Bayes Classifier 的局限性：
- 对二进制特征的限制：BNB 不适用于特征非二元的任务。如果特征具有两个以上的值，BNB 不会考虑该信息。
- 朴素假设：与其他朴素贝叶斯分类器一样，BNB 做出了特征独立性的假设，这对于某些任务来说可能过于简单。
伯努利朴素贝叶斯分类器对于具有二元特征的二元分类任务（例如文本的情感分析或垃圾邮件分类）是一个不错的选择。它易于使用，并且对于此类数据表现良好。
2.21.1.创建 Bernoulli Naive Bayes (BNB) Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Bernoulli Naive Bayes (BNB) Classifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_BNBClassifier.py # The code demonstrates the process of training Bernoulli Naive Bayes (BNB) Classifier on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.naive_bayes import BernoulliNB from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Bernoulli Naive Bayes (BNB) Classifier model bnb_model = BernoulliNB() # train the model on the entire dataset bnb_model.fit(X, y) # predict classes for the entire dataset y_pred = bnb_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Bernoulli Naive Bayes (BNB) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(bnb_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "bnb_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Bernoulli Naive Bayes (BNB) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Bernoulli Naive Bayes (BNB) Classifier model:0.3333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.33 1.00 0.50 50
Python 1 0.00 0.00 0.00 50
Python 2 0.00 0.00 0.00 50
Python
Python accuracy 0.33 150
Python macro avg 0.11 0.33 0.17 150
Python weighted avg 0.11 0.33 0.17 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\bnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Bernoulli Naive Bayes (BNB) Classifier model in ONNX format:0.3333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.33 1.00 0.50 50
Python 1 0.00 0.00 0.00 50
Python 2 0.00 0.00 0.00 50
Python
Python accuracy 0.33 150
Python macro avg 0.11 0.33 0.17 150
Python weighted avg 0.11 0.33 0.17 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\bnb_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Bernoulli Naive Bayes (BNB) Classifier model in ONNX format:0.3333333333333333
2.21.2.用于处理 Bernoulli Naive Bayes (BNB) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_BNBClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "bnb_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="BNBClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=51 FAILED [class=0, true class=1] features=(7.00,3.20,4.70,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=52 FAILED [class=0, true class=1] features=(6.40,3.20,4.50,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=53 FAILED [class=0, true class=1] features=(6.90,3.10,4.90,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=54 FAILED [class=0, true class=1] features=(5.50,2.30,4.00,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=55 FAILED [class=0, true class=1] features=(6.50,2.80,4.60,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=56 FAILED [class=0, true class=1] features=(5.70,2.80,4.50,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=57 FAILED [class=0, true class=1] features=(6.30,3.30,4.70,1.60] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=58 FAILED [class=0, true class=1] features=(4.90,2.40,3.30,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=59 FAILED [class=0, true class=1] features=(6.60,2.90,4.60,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=60 FAILED [class=0, true class=1] features=(5.20,2.70,3.90,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=61 FAILED [class=0, true class=1] features=(5.00,2.00,3.50,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=62 FAILED [class=0, true class=1] features=(5.90,3.00,4.20,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=63 FAILED [class=0, true class=1] features=(6.00,2.20,4.00,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=64 FAILED [class=0, true class=1] features=(6.10,2.90,4.70,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=65 FAILED [class=0, true class=1] features=(5.60,2.90,3.60,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=66 FAILED [class=0, true class=1] features=(6.70,3.10,4.40,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=67 FAILED [class=0, true class=1] features=(5.60,3.00,4.50,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=68 FAILED [class=0, true class=1] features=(5.80,2.70,4.10,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=69 FAILED [class=0, true class=1] features=(6.20,2.20,4.50,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=70 FAILED [class=0, true class=1] features=(5.60,2.50,3.90,1.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=71 FAILED [class=0, true class=1] features=(5.90,3.20,4.80,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=72 FAILED [class=0, true class=1] features=(6.10,2.80,4.00,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=73 FAILED [class=0, true class=1] features=(6.30,2.50,4.90,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=74 FAILED [class=0, true class=1] features=(6.10,2.80,4.70,1.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=75 FAILED [class=0, true class=1] features=(6.40,2.90,4.30,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=76 FAILED [class=0, true class=1] features=(6.60,3.00,4.40,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=77 FAILED [class=0, true class=1] features=(6.80,2.80,4.80,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=78 FAILED [class=0, true class=1] features=(6.70,3.00,5.00,1.70] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=79 FAILED [class=0, true class=1] features=(6.00,2.90,4.50,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=80 FAILED [class=0, true class=1] features=(5.70,2.60,3.50,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=81 FAILED [class=0, true class=1] features=(5.50,2.40,3.80,1.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=82 FAILED [class=0, true class=1] features=(5.50,2.40,3.70,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=83 FAILED [class=0, true class=1] features=(5.80,2.70,3.90,1.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=84 FAILED [class=0, true class=1] features=(6.00,2.70,5.10,1.60] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=85 FAILED [class=0, true class=1] features=(5.40,3.00,4.50,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=86 FAILED [class=0, true class=1] features=(6.00,3.40,4.50,1.60] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=87 FAILED [class=0, true class=1] features=(6.70,3.10,4.70,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=88 FAILED [class=0, true class=1] features=(6.30,2.30,4.40,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=89 FAILED [class=0, true class=1] features=(5.60,3.00,4.10,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=90 FAILED [class=0, true class=1] features=(5.50,2.50,4.00,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=91 FAILED [class=0, true class=1] features=(5.50,2.60,4.40,1.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=92 FAILED [class=0, true class=1] features=(6.10,3.00,4.60,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=93 FAILED [class=0, true class=1] features=(5.80,2.60,4.00,1.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=94 FAILED [class=0, true class=1] features=(5.00,2.30,3.30,1.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=95 FAILED [class=0, true class=1] features=(5.60,2.70,4.20,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=96 FAILED [class=0, true class=1] features=(5.70,3.00,4.20,1.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=97 FAILED [class=0, true class=1] features=(5.70,2.90,4.20,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=98 FAILED [class=0, true class=1] features=(6.20,2.90,4.30,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=99 FAILED [class=0, true class=1] features=(5.10,2.50,3.00,1.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=100 FAILED [class=0, true class=1] features=(5.70,2.80,4.10,1.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=101 FAILED [class=0, true class=2] features=(6.30,3.30,6.00,2.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=102 FAILED [class=0, true class=2] features=(5.80,2.70,5.10,1.90] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=103 FAILED [class=0, true class=2] features=(7.10,3.00,5.90,2.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=104 FAILED [class=0, true class=2] features=(6.30,2.90,5.60,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=105 FAILED [class=0, true class=2] features=(6.50,3.00,5.80,2.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=106 FAILED [class=0, true class=2] features=(7.60,3.00,6.60,2.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=107 FAILED [class=0, true class=2] features=(4.90,2.50,4.50,1.70] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=108 FAILED [class=0, true class=2] features=(7.30,2.90,6.30,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=109 FAILED [class=0, true class=2] features=(6.70,2.50,5.80,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=110 FAILED [class=0, true class=2] features=(7.20,3.60,6.10,2.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=111 FAILED [class=0, true class=2] features=(6.50,3.20,5.10,2.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=112 FAILED [class=0, true class=2] features=(6.40,2.70,5.30,1.90] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=113 FAILED [class=0, true class=2] features=(6.80,3.00,5.50,2.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=114 FAILED [class=0, true class=2] features=(5.70,2.50,5.00,2.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=115 FAILED [class=0, true class=2] features=(5.80,2.80,5.10,2.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=116 FAILED [class=0, true class=2] features=(6.40,3.20,5.30,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=117 FAILED [class=0, true class=2] features=(6.50,3.00,5.50,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=118 FAILED [class=0, true class=2] features=(7.70,3.80,6.70,2.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=119 FAILED [class=0, true class=2] features=(7.70,2.60,6.90,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=120 FAILED [class=0, true class=2] features=(6.00,2.20,5.00,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=121 FAILED [class=0, true class=2] features=(6.90,3.20,5.70,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=122 FAILED [class=0, true class=2] features=(5.60,2.80,4.90,2.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=123 FAILED [class=0, true class=2] features=(7.70,2.80,6.70,2.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=124 FAILED [class=0, true class=2] features=(6.30,2.70,4.90,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=125 FAILED [class=0, true class=2] features=(6.70,3.30,5.70,2.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=126 FAILED [class=0, true class=2] features=(7.20,3.20,6.00,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=127 FAILED [class=0, true class=2] features=(6.20,2.80,4.80,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=128 FAILED [class=0, true class=2] features=(6.10,3.00,4.90,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=129 FAILED [class=0, true class=2] features=(6.40,2.80,5.60,2.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=130 FAILED [class=0, true class=2] features=(7.20,3.00,5.80,1.60] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=131 FAILED [class=0, true class=2] features=(7.40,2.80,6.10,1.90] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=132 FAILED [class=0, true class=2] features=(7.90,3.80,6.40,2.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=133 FAILED [class=0, true class=2] features=(6.40,2.80,5.60,2.20] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=134 FAILED [class=0, true class=2] features=(6.30,2.80,5.10,1.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=135 FAILED [class=0, true class=2] features=(6.10,2.60,5.60,1.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=136 FAILED [class=0, true class=2] features=(7.70,3.00,6.10,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=137 FAILED [class=0, true class=2] features=(6.30,3.40,5.60,2.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=138 FAILED [class=0, true class=2] features=(6.40,3.10,5.50,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=139 FAILED [class=0, true class=2] features=(6.00,3.00,4.80,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=140 FAILED [class=0, true class=2] features=(6.90,3.10,5.40,2.10] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=141 FAILED [class=0, true class=2] features=(6.70,3.10,5.60,2.40] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=142 FAILED [class=0, true class=2] features=(6.90,3.10,5.10,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=143 FAILED [class=0, true class=2] features=(5.80,2.70,5.10,1.90] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=144 FAILED [class=0, true class=2] features=(6.80,3.20,5.90,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=145 FAILED [class=0, true class=2] features=(6.70,3.30,5.70,2.50] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=146 FAILED [class=0, true class=2] features=(6.70,3.00,5.20,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=147 FAILED [class=0, true class=2] features=(6.30,2.50,5.00,1.90] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=148 FAILED [class=0, true class=2] features=(6.50,3.00,5.20,2.00] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=149 FAILED [class=0, true class=2] features=(6.20,3.40,5.40,2.30] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier sample=150 FAILED [class=0, true class=2] features=(5.90,3.00,5.10,1.80] Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier correct results: 33.33% Iris_BNBClassifier (EURUSD,H1) model=BNBClassifier all samples accuracy=0.333333 Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=1] features=(6.30,2.50,4.90,1.50) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=2] features=(6.30,2.70,4.90,1.80) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=1] features=(7.00,3.20,4.70,1.40) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=1] features=(6.40,3.20,4.50,1.50) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=2] features=(6.30,3.30,6.00,2.50) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=2] features=(5.80,2.70,5.10,1.90) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=2] features=(7.10,3.00,5.90,2.10) Iris_BNBClassifier (EURUSD,H1) model:BNBClassifier FAILED [class=0, true class=2] features=(6.30,2.90,5.60,1.80) Iris_BNBClassifier (EURUSD,H1) model=BNBClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 33.33%，与原始模型的准确率相对应。
2.21.3.Bernoulli Naive Bayes (BNB) Classifier 的 ONNX 表示
图 35.Netron 中 Bernoulli Naive Bayes (BNB) Classifier 的 ONNX 表示
2.22.Multilayer Perceptron Classifier
Multilayer Perceptron (MLP，多层感知器) Classifier 是一个用于分类任务的多层神经网络。它由多层神经元组成，包括输入层、隐藏层和输出层。MLP Classifier 能够学习数据中复杂的非线性依赖关系。
MLP Classifier 的原理：
- 多层架构：MLP Classifier 具有多层架构，包括输入层、一个或多个隐藏层和输出层。层中的每个神经元都通过学习到的权重与相邻层中的神经元相连。
- 激活函数：在每个神经元内部，都应用一个激活函数，将非线性引入模型，并允许 MLP Classifier 对复杂的数据依赖关系进行建模。
- 使用反向传播进行训练：MLP Classifier 使用反向传播方法进行训练，最大限度地减少模型预测和真实类标签之间的误差。
MLP Classifier 的优点：
- 建模复杂依赖关系的能力：MLP Classifier 可以学习数据中复杂的非线性依赖关系，因此可以在简单线性模型不足的任务中表现良好。
- 多功能性：MLP Classifier 可用于广泛的分类任务，包括多类分类和多任务问题。
MLP Classifier 的局限性：
- 对超参数的敏感性：MLP Classifier 有许多超参数，例如隐藏层的数量、每层的神经元数量、学习率等。调整这些参数可能非常耗时且耗资源。
- 大量数据的要求：MLP Classifier 需要大量的训练数据以避免过度拟合，尤其是当模型具有许多参数时。
- 过度拟合：如果模型参数太多或数据不足，则可能会过度拟合并在新数据上表现不佳。
MLP Classifier 是用于分类任务的强大工具，尤其是当数据表现出复杂的依赖关系时。它通常用于机器学习和深度学习领域解决各种分类问题。然而，适当调整超参数并确保足够数量的训练数据对于成功应用该模型至关重要。
2.22.1.创建 Multilayer Perceptron Classifier 模型的代码
该代码演示了在 Iris 数据集上训练 Multilayer Perceptron Classifier、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_MLPClassifier.py # The code demonstrates the process of training Multilayer Perceptron Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.neural_network import MLPClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Multilayer Perceptron (MLP) Classifier model mlp_model = MLPClassifier(max_iter=1000, random_state=42) # train the model on the entire dataset mlp_model.fit(X, y) # predict classes for the entire dataset y_pred = mlp_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Multilayer Perceptron (MLP) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(mlp_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path +"mlp_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Multilayer Perceptron (MLP) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Multilayer Perceptron (MLP) Classifier model:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 0.94 0.97 50
Python 2 0.94 1.00 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\mlp_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Multilayer Perceptron (MLP) Classifier model in ONNX format:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 0.94 0.97 50
Python 2 0.94 1.00 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\mlp_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Multilayer Perceptron (MLP) Classifier model in ONNX format:0.98
2.22.2.用于处理 Multilayer Perceptron Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_MLPClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "mlp_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="MLPClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_MLPClassifier (EURUSD,H1) model:MLPClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_MLPClassifier (EURUSD,H1) model:MLPClassifier sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_MLPClassifier (EURUSD,H1) model:MLPClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_MLPClassifier (EURUSD,H1) model:MLPClassifier correct results: 98.00% Iris_MLPClassifier (EURUSD,H1) model=MLPClassifier all samples accuracy=0.980000 Iris_MLPClassifier (EURUSD,H1) model:MLPClassifier FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50) Iris_MLPClassifier (EURUSD,H1) model=MLPClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 98%，与原始模型的准确率一致。
2.22.3.Multilayer Perceptron Classifier 的 ONNX 表示
图 36.Netron 中 Multilayer Perceptron Classifier 的 ONNX 表示
2.23.Linear Discriminant Analysis (LDA) Classifier
Linear Discriminant Analysis（LDA，线性判别分析）Classifier 是一种用于分类任务的机器学习方法。它属于降维方法和低维空间分类的范畴。LDA 构建超平面来最大化类别之间的分离。
LDA Classifier 的原理：
- 降维：LDA的核心思想是降维。其目的是找到一个数据类别最大程度分离的新特征空间。
- 最大化分离：LDA 构建超平面（特征的线性组合），最大化不同类别中特征的平均值之间的差异，并最小化每个类别内的方差。
- 训练参数：LDA模型在训练数据集上进行训练，计算超平面的参数和数据投影到新的特征空间中。
LDA Classifier 的优点：
- 改进了类别分离：LDA 可以显著改善数据的类别分离，特别是在原始特征空间中类别严重重叠的情况下。
- 降维：LDA 还可以用于数据降维，这对于可视化和降低计算复杂性很有用。
LDA Classifier 的局限性：
- 正态性假设：LDA 假设特征遵循正态分布，并且类别具有相等的协方差矩阵。如果不满足这些假设，LDA 可能会提供不太准确的结果。
- 对异常值的敏感性：LDA 对数据中的异常值很敏感，因为它们会影响模型参数计算。
- 多类分类中的挑战：LDA 最初是为二元分类而开发的，其向多类任务的扩展需要进行调整。
LDA Classifier 对于分类和降维任务来说是一种很有价值的方法，尤其是在需要改善类别分离时。它经常用于统计学、生物学、医学分析等领域的数据分析和分类。
2.23.1.创建 Linear Discriminant Analysis (LDA) Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Linear Discriminant Analysis (LDA) Classifier、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_LDAClassifier.py # The code demonstrates the process of training Linear Discriminant Analysis (LDA) Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.discriminant_analysis import LinearDiscriminantAnalysis from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Linear Discriminant Analysis (LDA) Classifier model lda_model = LinearDiscriminantAnalysis() # train the model on the entire dataset lda_model.fit(X, y) # predict classes for the entire dataset y_pred = lda_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Linear Discriminant Analysis (LDA) Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(lda_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path +"lda_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Linear Discriminant Analysis (LDA) Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Linear Discriminant Analysis (LDA) Classifier model:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.96 0.97 50
Python 2 0.96 0.98 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\lda_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Linear Discriminant Analysis (LDA) Classifier model in ONNX format:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.96 0.97 50
Python 2 0.96 0.98 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\lda_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Linear Discriminant Analysis (LDA) Classifier model in ONNX format:0.98
2.23.2.用于处理 Linear Discriminant Analysis (LDA) Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_LDAClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "lda_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="LDAClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_LDAClassifier (EURUSD,H1) model:LDAClassifier sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_LDAClassifier (EURUSD,H1) model:LDAClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_LDAClassifier (EURUSD,H1) model:LDAClassifier sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_LDAClassifier (EURUSD,H1) model:LDAClassifier correct results: 98.00% Iris_LDAClassifier (EURUSD,H1) model=LDAClassifier all samples accuracy=0.980000 Iris_LDAClassifier (EURUSD,H1) model=LDAClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 98%，与原始模型的准确率一致。
2.23.3.Linear Discriminant Analysis (LDA) Classifier 的 ONNX 表示
图 37.Netron 中 Linear Discriminant Analysis (LDA) Classifier 的 ONNX 表示
2.24.Hist Gradient Boosting
Hist Gradient Boosting Classifier 是一种属于梯度提升家族的机器学习算法，专为分类任务而设计。它是一种广泛应用于数据分析和机器学习的高效而强大的方法。
Hist Gradient Boosting Classifier 的原理：
- 梯度提升：Hist Gradient Boosting Classifier 基于梯度提升方法，它构建决策树集合来改进分类。它通过连续训练弱模型并纠正先前模型的错误来实现这一点。
- 直方图用法：名称中的“Hist”表示该算法使用直方图进行有效的数据处理。Hist Gradient Boosting 无需进行详尽的特征枚举，而是构建特征直方图，从而可以快速构建决策树。
- 残差训练：与其他梯度提升方法一样，Hist Gradient Boosting 在前一个模型的残差上训练每棵新树，以优化预测。
Hist Gradient Boosting Classifier 的优点：
- 高准确率：Hist Gradient Boosting Classifier 通常提供较高的分类准确度，尤其是在使用大量树时。
- 效率：使用直方图可以使算法有效地处理大型数据集并快速构建集成。
- 处理异构数据的能力：该算法可以处理异构数据，包括分类和数字特征。
Hist Gradient Boosting Classifier 的局限性：
- 对过度拟合的敏感性：当参数调整不当或使用大量树时，Hist Gradient Boosting Classifier 可能容易过度拟合。
- 参数调整：与其他梯度提升算法一样，Hist Gradient Boosting 需要仔细调整参数才能获得最佳性能。
Hist Gradient Boosting Classifier 是一种用于分类和回归任务的强大算法，可在数据处理中提供高精度和高效率。它可以应用于数据分析、生物信息学、金融等各个领域。
2.24.1.创建 Histogram-Based Gradient Boosting Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 Hist Gradient Boosting Classifier、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_HistGradientBoostingClassifier.py # The code demonstrates the process of training Histogram-Based Gradient Boosting Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.experimental import enable_hist_gradient_boosting from sklearn.ensemble import HistGradientBoostingClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a Histogram-Based Gradient Boosting Classifier model hist_gradient_boosting_model = HistGradientBoostingClassifier(random_state=42) # train the model on the entire dataset hist_gradient_boosting_model.fit(X, y) # predict classes for the entire dataset y_pred = hist_gradient_boosting_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of Hist Gradient Boosting Classifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(hist_gradient_boosting_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path +"hist_gradient_boosting_classifier_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of Hist Gradient Boosting Classifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Hist Gradient Boosting Classifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\hist_gradient_boosting_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Hist Gradient Boosting Classifier model in ONNX format:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\hist_gradient_boosting_classifier_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of Hist Gradient Boosting Classifier model in ONNX format:1.0
2.24.2.用于处理 Hist Gradient Boosting Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_HistGradientBoostingClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "hist_gradient_boosting_classifier_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="HistGradientBoostingClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_HistGradientBoostingClassifier (EURUSD,H1) model:HistGradientBoostingClassifier correct results: 100.00% Iris_HistGradientBoostingClassifier (EURUSD,H1) model=HistGradientBoostingClassifier all samples accuracy=1.000000 Iris_HistGradientBoostingClassifier (EURUSD,H1) model=HistGradientBoostingClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率一致。
2.24.3.Hist Gradient Boosting Classifier 的 ONNX 表示
图 38.Netron 中 Hist Gradient Boosting Classifier 的 ONNX 表示
2.25。CategoricalNB Classifier
CategoricalNB 是一种基于贝叶斯定理的分类算法。它专为具有分类特征的数据集而设计，广泛应用于文本分类、垃圾邮件检测和其他涉及离散数据的应用。
CategoricalNB 的原理：
- 朴素贝叶斯分类器：CategoricalNB 是一种基于贝叶斯定理的朴素贝叶斯分类器。它使用给定类别的每个特征的条件概率来计算一组特征属于特定类别的概率。
- 类别特征：与假设连续特征服从正态分布的 Gaussian Naive Bayes Classifier 不同，CategoricalNB 适用于具有分类特征的数据集。它为每个类别的每个特征建模概率分布。
- 独立性假设：Naive Bayes Classifier中的“naive”来自于特征独立的假设。CategoricalNB 假设特征在给定类别的情况下是条件独立的。尽管在实践中并不总是满足这一假设，但朴素贝叶斯方法可以在许多现实世界的数据集上表现良好。
CategoricalNB 的优点：
- 效率：CategoricalNB 计算效率高，并且可扩展到大型数据集。它需要最少的内存并可以提供快速的预测。
- 可解释性：它的概率性质使得 CategoricalNB 可解释。它可以提供哪些特征影响预测的见解。
- 处理分类数据：CategoricalNB 是专为具有分类特征的数据集而设计的。它可以有效地处理文本数据和其他离散特征类型。
- 基准表现：它通常作为文本分类任务的强大基线模型，并且在小数据集上的表现优于更复杂的算法。
CategoricalNB 的局限性：
- 独立性假设：特征独立性的假设可能并不适用于所有数据集。如果特征高度依赖，CategoricalNB 的性能可能会下降。
- 对特征缩放的敏感度：CategoricalNB 不需要特征缩放，因为它适用于分类数据。然而，在某些情况下，以不同的方式规范化或编码分类特征可能会影响其性能。
- 表达能力有限：CategoricalNB 可能无法捕捉复杂的数据依赖关系以及更复杂的算法，例如深度学习模型。
- 处理缺失数据：它假设数据集中没有缺失值，缺失值需要进行预处理。
CategoricalNB 是一种有价值的分类算法，特别适合具有分类特征的数据集。它的简单性、效率和可解释性使其成为各种分类任务的有用工具。尽管存在独立性假设等局限性，它仍然是文本分类和其他离散数据占主导地位的任务的热门选择。处理分类数据时，将 CategoricalNB 视为基准模型通常是一个合理的选择。然而，与更复杂的模型相比，评估其性能非常重要，特别是当数据中存在特征依赖关系时。
2.25.1.创建 CategoricalNB Classifier 模型的代码
此代码演示了在 Iris 数据集上训练 CategoricalNB Classifier、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_CategoricalNBClassifier.py # The code demonstrates the process of training CategoricalNB Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.naive_bayes import CategoricalNB from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create a CategoricalNB model categorical_nb_model = CategoricalNB() # train the model on the entire dataset categorical_nb_model.fit(X, y) # predict classes for the entire dataset y_pred = categorical_nb_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of CategoricalNB model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(categorical_nb_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "categorical_nb_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of CategoricalNB model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of CategoricalNB model:0.9333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.86 0.96 0.91 50
Python 2 0.95 0.84 0.89 50
Python
Python accuracy 0.93 150
Python macro avg 0.94 0.93 0.93 150
Python weighted avg 0.94 0.93 0.93 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\categorical_nb_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of CategoricalNB model in ONNX format:0.9333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.86 0.96 0.91 50
Python 2 0.95 0.84 0.89 50
Python
Python accuracy 0.93 150
Python macro avg 0.94 0.93 0.93 150
Python weighted avg 0.94 0.93 0.93 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\categorical_nb_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of CategoricalNB model in ONNX format:0.9333333333333333
2.25.2.用于处理 CategoricalNB Classifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_CategoricalNBClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "categorical_nb_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="CategoricalNBClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+输出：
Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=102 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=122 FAILED [class=1, true class=2] features=(5.60,2.80,4.90,2.00] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=124 FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=128 FAILED [class=1, true class=2] features=(6.10,3.00,4.90,1.80] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier sample=143 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier correct results: 93.33% Iris_CategoricalNBClassifier (EURUSD,H1) model=CategoricalNBClassifier all samples accuracy=0.933333 Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80) Iris_CategoricalNBClassifier (EURUSD,H1) model:CategoricalNBClassifier FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90) Iris_CategoricalNBClassifier (EURUSD,H1) model=CategoricalNBClassifier batch test accuracy=0.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 93.33%，与原始模型的准确率一致。
2.25.3.CategoricalNB Classifier 的 ONNX 表示
图 39.Netron 中 CategoricalNB Classifier 的 ONNX 表示
关于 ExtraTreeClassifier 和 ExtraTreesClassifier 模型的注意事项
ExtraTreeClassifier 和 ExtraTreesClassifier 是两个不同的分类器，它们的主要区别在于它们的工作原理：
ExtraTreeClassifier（Extremely Randomized Trees Classifier，极端随机树分类器）：
- 该分类器也称为 Extremely Randomized Trees（极端随机树）或 Extra-Trees （额外树）。
- 它基于随机决策树的思想。
- 在 ExtraTreeClassifier 中，每个树节点的分割选择都是随机发生的，无需事先搜索最佳分割。
- 这使得分类器的计算强度低于经典随机森林，因为它不需要计算每个节点的最佳分割。
- ExtraTreeClassifier 通常对特征使用随机阈值并进行随机分割，从而产生更多随机树。
- 由于没有搜索最佳分割，ExtraTreeClassifier 与随机森林相比速度更快，但准确性较低。
ExtraTreesClassifier（Extremely Randomized Trees Classifier，极端随机树分类器）：
- ExtraTreesClassifier 也是一个基于极端随机树方法的分类器。
- ExtraTreesClassifier 和 ExtraTreeClassifier 之间的主要区别在于 ExtraTreesClassifier 执行随机分割以在每个树节点选择最佳分割。
- 这意味着 ExtraTreesClassifier 在选择最佳分割时应用具有额外随机性的随机森林。
- ExtraTreesClassifier 通常比 ExtraTreeClassifier 更准确，因为它执行随机分割以找到最佳分割特征。
- 然而，由于需要进行更广泛的最佳分割搜索，ExtraTreesClassifier 的计算量可能会更大。
综上所述，这两个分类器的主要区别在于分割选择的随机性程度。ExtraTreeClassifier 对每个节点做出随机选择而不搜索最佳分割，而 ExtraTreesClassifier 在每个节点上寻找最佳分割时执行随机分割。
2.26.ExtraTreeClassifier
ExtraTreeClassifier 的原理：
- 随机节点分割：ExtraTreeClassifier 的主要原理是它随机选择每个树节点的分割。这与选择最佳特征进行分割的传统决策树不同。ExtraTreeClassifier 在执行分割时不考虑最佳分割，使其更加随机且不易过度拟合。
- 结果汇总：在集成构建期间，ExtraTreeClassifier 创建多个随机树并汇总其结果。这样做是为了提高模型的泛化能力并减少方差。树集合有助于对抗过度拟合并提高预测稳定性。
- 随机阈值：在分割节点时，ExtraTreeClassifier 为每个特征选择随机阈值，而不是特定的最优值。这引入了更多的随机性和模型稳定性。
- 抗过度拟合：由于随机分割且没有最佳分割选择，与常规决策树相比，ExtraTreeClassifier 通常不太容易过度拟合。
- 训练速度快：与许多其他算法（例如随机森林）相比，ExtraTreeClassifier 训练所需的计算资源更少。这使得它能够快速、高效地处理大型数据集。
- 多功能性：ExtraTreeClassifier 可用于分类和回归任务，使其成为针对各种类型问题的多功能算法。
- 随机性：在某些情况下，使用随机分割可能会导致模型不太准确。仔细的参数调整非常重要。
- 对异常值的敏感性：ExtraTreeClassifier 在建立随机分割时对数据中的异常值很敏感。在某些情况下，这可能会导致预测不稳定。
- 可解释性较低：与常规决策树相比，ExtraTreeClassifier 的可解释性较差，也更难解释。
2.26.1.创建 ExtraTreeClassifier 模型的代码
此代码演示了在 Iris 数据集上训练 ExtraTreeClassifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型进行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_ExtraTreeClassifier.py # The code demonstrates the process of training ExtraTree Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.tree import ExtraTreeClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create an ExtraTreeClassifier model extra_tree_model = ExtraTreeClassifier() # train the model on the entire dataset extra_tree_model.fit(X, y) # predict classes for the entire dataset y_pred = extra_tree_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of ExtraTreeClassifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(extra_tree_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "extra_tree_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of ExtraTreeClassifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of ExtraTreeClassifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\extra_tree_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of ExtraTreeClassifier model in ONNX format:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\extra_tree_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of ExtraTreeClassifier model in ONNX format:1.0
2.26.2.用于处理 ExtraTreeClassifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_ExtraTreeClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "extra_tree_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="ExtraTreeClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_ExtraTreeClassifier (EURUSD,H1) model:ExtraTreeClassifier correct results: 100.00% Iris_ExtraTreeClassifier (EURUSD,H1) model=ExtraTreeClassifier all samples accuracy=1.000000 Iris_ExtraTreeClassifier (EURUSD,H1) model=ExtraTreeClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率一致。
2.26.3.ExtraTreeClassifier 的 ONNX 表示
图 40.Netron 中 ExtraTreeClassifier 的 ONNX 表示
2.27.ExtraTreesClassifier
ExtraTreesClassifier 是一种用于分类任务的强大的机器学习算法。该算法是随机森林的扩展和改进，具有一些优点和缺点。ExtraTreesClassifier 的原理：
- 自助抽样：与随机森林类似，ExtraTreesClassifier 使用自助方法从训练数据集中创建多个子样本。这意味着对于每棵树，都会从原始数据中创建一个可替换的随机子样本。
- 随机分割：与随机森林不同，随机森林为每个树节点选择最佳分割特征，而 ExtraTreesClassifier 使用随机特征和随机阈值来分割节点。这使得树更加随机并减少过度拟合。
- 投票：构建一组树之后，每棵树都会对对象的类别进行投票。最终，获得最多票数的类别成为预测类别。
- 减少过度拟合：与传统决策树相比，随机分割和随机特征的使用使得 ExtraTreesClassifier 不太容易过度拟合。
- 训练速度快：与其他一些算法（例如梯度提升）相比，ExtraTreesClassifier 训练所需的计算资源更少。这使得它快速而高效，特别是对于大型数据集。
- 异常值稳健性：由于树木和随机分割的集合，ExtraTreesClassifier 通常对数据中的异常值更具鲁棒性。
- 复杂的可解释性：由于随机分割和特征的数量众多，分析和解释 ExtraTreesClassifier 模型可能具有挑战性。
- 参数调整：尽管 ExtraTreesClassifier 非常高效，但它可能需要仔细调整超参数才能实现最佳性能。
- 并非总是表现最佳：在某些任务中，ExtraTreesClassifier 可能不如其他算法准确，例如梯度提升。
2.27.1.创建 ExtraTreesClassifier 模型的代码
此代码演示了在 Iris 数据集上训练 ExtraTreesClassifier 模型、将其导出为 ONNX 格式以及使用 ONNX 模型执行分类的过程。它还评估了原始模型和 ONNX 模型的准确性。
# Iris_ExtraTreesClassifier.py # The code demonstrates the process of training ExtraTrees Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original model and the ONNX model. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.ensemble import ExtraTreesClassifier from sklearn.metrics import accuracy_score, classification_report from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create an ExtraTreesClassifier model extra_trees_model = ExtraTreesClassifier() # train the model on the entire dataset extra_trees_model.fit(X, y) # predict classes for the entire dataset y_pred = extra_trees_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) print("Accuracy of ExtraTreesClassifier model:", accuracy) # display the classification report print("\nClassification Report:\n", classification_report(y, y_pred)) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(extra_trees_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + "extra_trees_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # print model path print(f"Model saved to {onnx_filename}") # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # display information about input tensors in ONNX print("\nInformation about input tensors in ONNX:") for i, input_tensor in enumerate(onnx_session.get_inputs()): print(f"{i + 1}. Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}") # display information about output tensors in ONNX print("\nInformation about output tensors in ONNX:") for i, output_tensor in enumerate(onnx_session.get_outputs()): print(f"{i + 1}. Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}") # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) print("\nAccuracy of ExtraTreesClassifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of ExtraTreesClassifier model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\extra_trees_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of ExtraTreesClassifier model in ONNX format:1.
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\extra_trees_iris.onnx
Python
Python Information about input tensors in ONNX:
Python 1.Name: float_input, Data Type: tensor(float), Shape: [None, 4]
Python
Python Information about output tensors in ONNX:
Python 1.Name: output_label, Data Type: tensor(int64), Shape: [None]
Python 2.Name: output_probability, Data Type: seq(map(int64,tensor(float))), Shape: []
Python
Python Accuracy of ExtraTreesClassifier model in ONNX format:1.
2.27.2.用于处理 ExtraTreesClassifier 模型的 MQL5 代码
//+------------------------------------------------------------------+ //| Iris_ExtraTreesClassifier.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" #resource "extra_trees_iris.onnx" as const uchar ExtModel[]; //+------------------------------------------------------------------+ //| Test IRIS dataset samples | //+------------------------------------------------------------------+ bool TestSamples(long model,float &input_data[][4], int &model_classes_id[]) { //--- check number of input samples ulong batch_size=input_data.Range(0); if(batch_size==0) return(false); //--- prepare output array ArrayResize(model_classes_id,(int)batch_size); //--- float output_data[]; //--- struct Map { ulong key[]; float value[]; } output_data_map[]; //--- check consistency bool res=ArrayResize(output_data,(int)batch_size)==batch_size; //--- if(res) { //--- set input shape ulong input_shape[]= {batch_size,input_data.Range(1)}; OnnxSetInputShape(model,0,input_shape); //--- set output shapeы ulong output_shape1[]= {batch_size}; ulong output_shape2[]= {batch_size}; OnnxSetOutputShape(model,0,output_shape1); OnnxSetOutputShape(model,1,output_shape2); //--- run the model res=OnnxRun(model,0,input_data,output_data,output_data_map); //--- postprocessing if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- for(uint n=0; n<output_data_map.Size(); n++) { int model_class_id=-1; int max_idx=-1; float max_value=-1; //--- copy to arrays ArrayCopy(output_keys,output_data_map[n].key); ArrayCopy(output_values,output_data_map[n].value); //ArrayPrint(output_keys); //ArrayPrint(output_values); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } //--- store the result to the output array model_classes_id[n]=model_class_id; //Print("model_class_id=",model_class_id); } } } //--- return(res); } //+------------------------------------------------------------------+ //| Test all samples from IRIS dataset (150) | //| Here we test all samples with batch=1, sample by sample | //+------------------------------------------------------------------+ bool TestAllIrisDataset(const long model,const string model_name,double &model_accuracy) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("iris dataset not prepared"); return(false); } //--- show dataset for(int k=0; k<total_samples; k++) { //PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); } //--- array for output classes int model_output_classes_id[]; //--- check all Iris dataset samples int correct_results=0; for(int k=0; k<total_samples; k++) { //--- input array float iris_sample_input_data[1][4]; //--- prepare input data from kth iris sample dataset iris_sample_input_data[0][0]=(float)iris_samples[k].features[0]; iris_sample_input_data[0][1]=(float)iris_samples[k].features[1]; iris_sample_input_data[0][2]=(float)iris_samples[k].features[2]; iris_sample_input_data[0][3]=(float)iris_samples[k].features[3]; //--- run model bool res=TestSamples(model,iris_sample_input_data,model_output_classes_id); //--- check result if(res) { if(model_output_classes_id[0]==iris_samples[k].class_id) { correct_results++; } else { PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_output_classes_id[0],iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } } model_accuracy=1.0*correct_results/total_samples; //--- PrintFormat("model:%s correct results: %.2f%%",model_name,100*model_accuracy); //--- return(true); } //+------------------------------------------------------------------+ //| Here we test batch execution of the model | //+------------------------------------------------------------------+ bool TestBatchExecution(const long model,const string model_name,double &model_accuracy) { model_accuracy=0; //--- array for output classes int model_output_classes_id[]; int correct_results=0; int total_results=0; bool res=false; //--- run batch with 3 samples float input_data_batch3[3][4]= { {5.1f,3.5f,1.4f,0.2f}, // iris dataset sample id=1, Iris-setosa {6.3f,2.5f,4.9f,1.5f}, // iris dataset sample id=73, Iris-versicolor {6.3f,2.7f,4.9f,1.8f} // iris dataset sample id=124, Iris-virginica }; int correct_classes_batch3[3]= {0,1,2}; //--- run model res=TestSamples(model,input_data_batch3,model_output_classes_id); if(res) { //--- check result for(int j=0; j<ArraySize(model_output_classes_id); j++) { //--- check result if(model_output_classes_id[j]==correct_classes_batch3[j]) correct_results++; else { PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch3[j],input_data_batch3[j][0],input_data_batch3[j][1],input_data_batch3[j][2],input_data_batch3[j][3]); } total_results++; } } else return(false); //--- run batch with 10 samples float input_data_batch10[10][4]= { {5.5f,3.5f,1.3f,0.2f}, // iris dataset sample id=37 (Iris-setosa) {4.9f,3.1f,1.5f,0.1f}, // iris dataset sample id=38 (Iris-setosa) {4.4f,3.0f,1.3f,0.2f}, // iris dataset sample id=39 (Iris-setosa) {5.0f,3.3f,1.4f,0.2f}, // iris dataset sample id=50 (Iris-setosa) {7.0f,3.2f,4.7f,1.4f}, // iris dataset sample id=51 (Iris-versicolor) {6.4f,3.2f,4.5f,1.5f}, // iris dataset sample id=52 (Iris-versicolor) {6.3f,3.3f,6.0f,2.5f}, // iris dataset sample id=101 (Iris-virginica) {5.8f,2.7f,5.1f,1.9f}, // iris dataset sample id=102 (Iris-virginica) {7.1f,3.0f,5.9f,2.1f}, // iris dataset sample id=103 (Iris-virginica) {6.3f,2.9f,5.6f,1.8f} // iris dataset sample id=104 (Iris-virginica) }; //--- correct classes for all 10 samples in the batch int correct_classes_batch10[10]= {0,0,0,0,1,1,2,2,2,2}; //--- run model res=TestSamples(model,input_data_batch10,model_output_classes_id); //--- check result if(res) { for(int j=0; j<ArraySize(model_output_classes_id); j++) { if(model_output_classes_id[j]==correct_classes_batch10[j]) correct_results++; else { double f1=input_data_batch10[j][0]; double f2=input_data_batch10[j][1]; double f3=input_data_batch10[j][2]; double f4=input_data_batch10[j][3]; PrintFormat("model:%s FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f)",model_name,model_output_classes_id[j],correct_classes_batch10[j],input_data_batch10[j][0],input_data_batch10[j][1],input_data_batch10[j][2],input_data_batch10[j][3]); } total_results++; } } else return(false); //--- calculate accuracy model_accuracy=correct_results/total_results; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { string model_name="ExtraTreesClassifier"; //--- long model=OnnxCreateFromBuffer(ExtModel,ONNX_DEFAULT); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- test all dataset double model_accuracy=0; //-- test sample by sample execution for all Iris dataset if(TestAllIrisDataset(model,model_name,model_accuracy)) PrintFormat("model=%s all samples accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- test batch execution for several samples if(TestBatchExecution(model,model_name,model_accuracy)) PrintFormat("model=%s batch test accuracy=%f",model_name,model_accuracy); else PrintFormat("error in testing model=%s ",model_name); //--- release model OnnxRelease(model); } return(0); } //+------------------------------------------------------------------+
输出：
Iris_ExtraTreesClassifier (EURUSD,H1) model:ExtraTreesClassifier correct results: 100.00% Iris_ExtraTreesClassifier (EURUSD,H1) model=ExtraTreesClassifier all samples accuracy=1.000000 Iris_ExtraTreesClassifier (EURUSD,H1) model=ExtraTreesClassifier batch test accuracy=1.000000
导出的 ONNX 模型在完整 Iris 数据集上的准确率为 100%，与原始模型的准确率一致。
2.27.3.ExtraTreesClassifier 的 ONNX 表示
图 41.Netron 中 ExtraTrees Classifier 的 ONNX 表示
2.28.比较所有模型的准确性
现在，让我们一起考虑所有模型并比较它们的性能。首先，我们将使用 Python 进行比较，然后在 MetaTrader 5 中加载并执行保存的 ONNX 模型。
2.28.1.计算所有模型并建立准确率比较图表的代码
该脚本在完整的 Fisher 鸢尾花数据集上计算来自 Scikit-learn 包的 27 个分类模型，将模型导出为 ONNX 格式，执行它们，并比较原始模型和 ONNX 模型的准确性。
# Iris_AllClassifiers.py # The code demonstrates the process of training 27 Classifier models on the Iris dataset, exports them to ONNX format, and making predictions using the ONNX model. # It also evaluates the accuracy of both the original and the ONNX models. # Copyright 2023, MetaQuotes Ltd. # https://www.mql5.com # import necessary libraries from sklearn import datasets from sklearn.metrics import accuracy_score from skl2onnx import convert_sklearn from skl2onnx.common.data_types import FloatTensorType import onnxruntime as ort import numpy as np import matplotlib.pyplot as plt from sys import argv # define the path for saving the model data_path = argv[0] last_index = data_path.rfind("\\") + 1 data_path = data_path[0:last_index] # load the Iris dataset iris = datasets.load_iris() X = iris.data y = iris.target # create and train each classifier model from sklearn.svm import SVC svc_model = SVC() svc_model.fit(X, y) from sklearn.ensemble import RandomForestClassifier random_forest_model = RandomForestClassifier(random_state=42) random_forest_model.fit(X, y) from sklearn.ensemble import GradientBoostingClassifier gradient_boosting_model = GradientBoostingClassifier(random_state=42) gradient_boosting_model.fit(X, y) from sklearn.ensemble import AdaBoostClassifier adaboost_model = AdaBoostClassifier(random_state=42) adaboost_model.fit(X, y) from sklearn.ensemble import BaggingClassifier bagging_model = BaggingClassifier(random_state=42) bagging_model.fit(X, y) from sklearn.neighbors import KNeighborsClassifier knn_model = KNeighborsClassifier() knn_model.fit(X, y) from sklearn.neighbors import RadiusNeighborsClassifier radius_neighbors_model = RadiusNeighborsClassifier(radius=1.0) radius_neighbors_model.fit(X, y) from sklearn.tree import DecisionTreeClassifier decision_tree_model = DecisionTreeClassifier(random_state=42) decision_tree_model.fit(X, y) from sklearn.linear_model import LogisticRegression logistic_regression_model = LogisticRegression(max_iter=1000, random_state=42) logistic_regression_model.fit(X, y) from sklearn.linear_model import RidgeClassifier ridge_classifier_model = RidgeClassifier(random_state=42) ridge_classifier_model.fit(X, y) from sklearn.linear_model import PassiveAggressiveClassifier passive_aggressive_model = PassiveAggressiveClassifier(max_iter=1000, random_state=42) passive_aggressive_model.fit(X, y) from sklearn.linear_model import Perceptron perceptron_model = Perceptron(max_iter=1000, random_state=42) perceptron_model.fit(X, y) from sklearn.linear_model import SGDClassifier sgd_model = SGDClassifier(max_iter=1000, random_state=42) sgd_model.fit(X, y) from sklearn.naive_bayes import GaussianNB gaussian_nb_model = GaussianNB() gaussian_nb_model.fit(X, y) from sklearn.naive_bayes import MultinomialNB multinomial_nb_model = MultinomialNB() multinomial_nb_model.fit(X, y) from sklearn.naive_bayes import ComplementNB complement_nb_model = ComplementNB() complement_nb_model.fit(X, y) from sklearn.naive_bayes import BernoulliNB bernoulli_nb_model = BernoulliNB() bernoulli_nb_model.fit(X, y) from sklearn.naive_bayes import CategoricalNB categorical_nb_model = CategoricalNB() categorical_nb_model.fit(X, y) from sklearn.tree import ExtraTreeClassifier extra_tree_model = ExtraTreeClassifier(random_state=42) extra_tree_model.fit(X, y) from sklearn.ensemble import ExtraTreesClassifier extra_trees_model = ExtraTreesClassifier(random_state=42) extra_trees_model.fit(X, y) from sklearn.svm import LinearSVC # Import LinearSVC linear_svc_model = LinearSVC(random_state=42) linear_svc_model.fit(X, y) from sklearn.svm import NuSVC nu_svc_model = NuSVC() nu_svc_model.fit(X, y) from sklearn.linear_model import LogisticRegressionCV logistic_regression_cv_model = LogisticRegressionCV(cv=5, max_iter=1000, random_state=42) logistic_regression_cv_model.fit(X, y) from sklearn.neural_network import MLPClassifier mlp_model = MLPClassifier(max_iter=1000, random_state=42) mlp_model.fit(X, y) from sklearn.discriminant_analysis import LinearDiscriminantAnalysis lda_model = LinearDiscriminantAnalysis() lda_model.fit(X, y) from sklearn.experimental import enable_hist_gradient_boosting from sklearn.ensemble import HistGradientBoostingClassifier hist_gradient_boosting_model = HistGradientBoostingClassifier(random_state=42) hist_gradient_boosting_model.fit(X, y) from sklearn.linear_model import RidgeClassifierCV ridge_classifier_cv_model = RidgeClassifierCV() ridge_classifier_cv_model.fit(X, y) # define a dictionary to store results results = {} # loop through the models for model_name, classifier_model in [ ('SVC Classifier', svc_model), ('Random Forest Classifier', random_forest_model), ('Gradient Boosting Classifier', gradient_boosting_model), ('AdaBoost Classifier', adaboost_model), ('Bagging Classifier', bagging_model), ('K-NN Classifier', knn_model), ('Radius Neighbors Classifier', radius_neighbors_model), ('Decision Tree Classifier', decision_tree_model), ('Logistic Regression Classifier', logistic_regression_model), ('Ridge Classifier', ridge_classifier_model), ('Ridge ClassifierCV', ridge_classifier_cv_model), ('Passive-Aggressive Classifier', passive_aggressive_model), ('Perceptron Classifier', perceptron_model), ('SGD Classifier', sgd_model), ('Gaussian Naive Bayes Classifier', gaussian_nb_model), ('Multinomial Naive Bayes Classifier', multinomial_nb_model), ('Complement Naive Bayes Classifier', complement_nb_model), ('Bernoulli Naive Bayes Classifier', bernoulli_nb_model), ('Categorical Naive Bayes Classifier', categorical_nb_model), ('Extra Tree Classifier', extra_tree_model), ('Extra Trees Classifier', extra_trees_model), ('LinearSVC Classifier', linear_svc_model), ('NuSVC Classifier', nu_svc_model), ('Logistic RegressionCV Classifier', logistic_regression_cv_model), ('MLP Classifier', mlp_model), ('Linear Discriminant Analysis Classifier', lda_model), ('Hist Gradient Boosting Classifier', hist_gradient_boosting_model) ]: # predict classes for the entire dataset y_pred = classifier_model.predict(X) # evaluate the model's accuracy accuracy = accuracy_score(y, y_pred) # define the input data type initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))] # export the model to ONNX format with float data type onnx_model = convert_sklearn(classifier_model, initial_types=initial_type, target_opset=12) # save the model to a file onnx_filename = data_path + f"{model_name.lower().replace(' ', '_')}_iris.onnx" with open(onnx_filename, "wb") as f: f.write(onnx_model.SerializeToString()) # load the ONNX model and make predictions onnx_session = ort.InferenceSession(onnx_filename) input_name = onnx_session.get_inputs()[0].name output_name = onnx_session.get_outputs()[0].name # convert data to floating-point format (float32) X_float32 = X.astype(np.float32) # predict classes for the entire dataset using ONNX y_pred_onnx = onnx_session.run([output_name], {input_name: X_float32})[0] # evaluate the accuracy of the ONNX model accuracy_onnx = accuracy_score(y, y_pred_onnx) # store results results[model_name] = { 'accuracy': accuracy, 'accuracy_onnx': accuracy_onnx } # print the accuracy of the original model and the ONNX model #print(f"{model_name} - Original Accuracy: {accuracy}, ONNX Accuracy: {accuracy_onnx}") # sort the models based on accuracy sorted_results = dict(sorted(results.items(), key=lambda item: item[1]['accuracy'], reverse=True)) # print the sorted results print("Sorted Results:") for model_name, metrics in sorted_results.items(): print(f"{model_name} - Original Accuracy: {metrics['accuracy']:.4f}, ONNX Accuracy: {metrics['accuracy_onnx']:.4f}") # create comparison plots for sorted results fig, ax = plt.subplots(figsize=(12, 8)) model_names = list(sorted_results.keys()) accuracies = [sorted_results[model_name]['accuracy'] for model_name in model_names] accuracies_onnx = [sorted_results[model_name]['accuracy_onnx'] for model_name in model_names] bar_width = 0.35 index = range(len(model_names)) bar1 = plt.bar(index, accuracies, bar_width, label='Model Accuracy') bar2 = plt.bar([i + bar_width for i in index], accuracies_onnx, bar_width, label='ONNX Accuracy') plt.xlabel('Models') plt.ylabel('Accuracy') plt.title('Comparison of Model and ONNX Accuracy (Sorted)') plt.xticks([i + bar_width / 2 for i in index], model_names, rotation=90, ha='center') plt.legend() plt.tight_layout() plt.show()
输出：
Python Sorted Results: Python Random Forest Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Gradient Boosting Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Bagging Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Decision Tree Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Extra Tree Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Extra Trees Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Hist Gradient Boosting Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 Python Logistic RegressionCV Classifier - Original Accuracy: 0.9800, ONNX Accuracy: 0.9800 Python MLP Classifier - Original Accuracy: 0.9800, ONNX Accuracy: 0.9800 Python Linear Discriminant Analysis Classifier - Original Accuracy: 0.9800, ONNX Accuracy: 0.9800 Python SVC Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 Python Radius Neighbors Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 Python Logistic Regression Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 Python NuSVC Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 Python K-NN Classifier - Original Accuracy: 0.9667, ONNX Accuracy: 0.9667 Python LinearSVC Classifier - Original Accuracy: 0.9667, ONNX Accuracy: 0.9667 Python AdaBoost Classifier - Original Accuracy: 0.9600, ONNX Accuracy: 0.9600 Python Passive-Aggressive Classifier - Original Accuracy: 0.9600, ONNX Accuracy: 0.9600 Python Gaussian Naive Bayes Classifier - Original Accuracy: 0.9600, ONNX Accuracy: 0.9600 Python Multinomial Naive Bayes Classifier - Original Accuracy: 0.9533, ONNX Accuracy: 0.9533 Python SGD Classifier - Original Accuracy: 0.9333, ONNX Accuracy: 0.9333 Python Categorical Naive Bayes Classifier - Original Accuracy: 0.9333, ONNX Accuracy: 0.9333 Python Ridge Classifier - Original Accuracy: 0.8533, ONNX Accuracy: 0.8533 Python Ridge ClassifierCV - Original Accuracy: 0.8533, ONNX Accuracy: 0.8533 Python Complement Naive Bayes Classifier - Original Accuracy: 0.6667, ONNX Accuracy: 0.6667 Python Perceptron Classifier - Original Accuracy: 0.6133, ONNX Accuracy: 0.6133 Python Bernoulli Naive Bayes Classifier - Original Accuracy: 0.3333, ONNX Accuracy: 0.3333该脚本还将生成一个包含所有 27 个模型的摘要结果的图像。
图 42.27 个分类模型及其 ONNX 版本在 Iris 数据集上的准确率比较
根据原始模型及其 ONNX 版本的准确率评估结果，可以得出以下结论：
七个模型在原始版本和 ONNX 版本中均表现出完美的准确率 (1.0000)。这些模型包括：
- Random Forest Classifier
- Gradient Boosting Classifier
- Bagging Classifier
- Decision Tree Classifier
- Extra Tree Classifier
- Extra Trees Classifier
- Hist Gradient Boosting Classifier
这些模型的 ONNX 格式也保持了高精度。
Logistic RegressionCV Classifier、MLP Classifier、Linear Discriminant Analysis Classifier 三个模型在原始版本和ONNX版本中都实现了高精度，精度达到0.9800。这些模型在两种格式中都表现良好。
包括 SVC Classifier、Radius Neighbors Classifier、NuSVC Classifier、K-NN Classifier、LinearSVC Classifier、AdaBoost Classifier、Passive-Aggressive Classifier、Gaussian Naive Bayes Classifier 和 Multinomial Naive Bayes Classifier 在内的几个模型在原始版本和 ONNX 版本中都表现出了良好的准确度，准确度得分分别为 0.9733、0.9667 或 0.9600。这些模型也在 ONNX 格式中保持了其准确性。
SGD Classifier、Categorical Naive Bayes Classifier、Ridge Classifier、Complement Naive Bayes Classifier、Perceptron Classifier 和 Bernoulli Naive Bayes Classifier 等模型的准确率较低，但在 ONNX 中保持准确率方面仍然表现良好。
所有考虑的模型在导出为 ONNX 格式时均能保持其准确性，这表明 ONNX 提供了一种保存和恢复机器学习模型的有效方法。但是，重要的是要记住，导出模型的质量取决于特定的模型算法和参数。
2.28.2.用于执行所有 ONNX 模型的 MQL5 代码
该脚本在完整的 Fisher 鸢尾花数据集上执行 2.28.1 中的脚本保存的所有 ONNX 模型。
//+------------------------------------------------------------------+ //| Iris_AllClassifiers.mq5 | //| Copyright 2023, MetaQuotes Ltd. | //| https://www.mql5.com | //+------------------------------------------------------------------+ #property copyright "Copyright 2023, MetaQuotes Ltd." #property link "https://www.mql5.com" #property version "1.00" #include "iris.mqh" //+------------------------------------------------------------------+ //| TestSampleSequenceMapOutput | //+------------------------------------------------------------------+ bool TestSampleSequenceMapOutput(long model,sIRISsample &iris_sample, int &model_class_id) { //--- model_class_id=-1; float input_data[1][4]; for(int k=0; k<4; k++) { input_data[0][k]=(float)iris_sample.features[k]; } //--- float out1[]; //--- struct Map { ulong key[]; float value[]; } out2[]; //--- bool res=ArrayResize(out1,input_data.Range(0))==input_data.Range(0); //--- if(res) { ulong input_shape[]= { input_data.Range(0), input_data.Range(1) }; ulong output_shape[]= { input_data.Range(0) }; //--- OnnxSetInputShape(model,0,input_shape); OnnxSetOutputShape(model,0,output_shape); //--- res=OnnxRun(model,0,input_data,out1,out2); //--- if(res) { //--- postprocessing of sequence map data //--- find class with maximum probability ulong output_keys[]; float output_values[]; //--- model_class_id=-1; int max_idx=-1; float max_value=-1; //--- for(uint n=0; n<out2.Size(); n++) { //--- copy to arrays ArrayCopy(output_keys,out2[n].key); ArrayCopy(output_values,out2[n].value); //--- find the key with maximum probability for(int k=0; k<ArraySize(output_values); k++) { if(k==0) { max_idx=0; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } else { if(output_values[k]>max_value) { max_idx=k; max_value=output_values[max_idx]; model_class_id=(int)output_keys[max_idx]; } } } } } } //--- return(res); } //+------------------------------------------------------------------+ //| TestSampleTensorOutput | //+------------------------------------------------------------------+ bool TestSampleTensorOutput(long model,sIRISsample &iris_sample, int &model_class_id) { //--- model_class_id=-1; float input_data[1][4]; for(int k=0; k<4; k++) { input_data[0][k]=(float)iris_sample.features[k]; } //--- ulong input_shape[]= { 1, 4}; OnnxSetInputShape(model,0,input_shape); //--- int output1[1]; float output2[1,3]; //--- ulong output_shape[]= {1}; OnnxSetOutputShape(model,0,output_shape); //--- ulong output_shape2[]= {1,3}; OnnxSetOutputShape(model,1,output_shape2); //--- bool res=OnnxRun(model,0,input_data,output1,output2); //--- class for these models in output1[0]; if(res) model_class_id=output1[0]; //--- return(res); } //+------------------------------------------------------------------+ //| Script program start function | //+------------------------------------------------------------------+ int OnStart(void) { sIRISsample iris_samples[]; //--- load dataset from file PrepareIrisDataset(iris_samples); //--- test int total_samples=ArraySize(iris_samples); if(total_samples==0) { Print("error in loading iris dataset from iris.csv"); return(false); } /*for(int k=0; k<total_samples; k++) { PrintFormat("%d (%.2f,%.2f,%.2f,%.2f) class %d (%s)",iris_samples[k].sample_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3],iris_samples[k].class_id,iris_samples[k].class_name); }*/ //---- string iris_models[]= { "random_forest_classifier_iris.onnx", "gradient_boosting_classifier_iris.onnx", "bagging_classifier_iris.onnx", "decision_tree_classifier_iris.onnx", "extra_tree_classifier_iris.onnx", "extra_trees_classifier_iris.onnx", "hist_gradient_boosting_classifier_iris.onnx", "logistic_regressioncv_classifier_iris.onnx", "mlp_classifier_iris.onnx", "linear_discriminant_analysis_classifier_iris.onnx", "svc_classifier_iris.onnx", "radius_neighbors_classifier_iris.onnx", "logistic_regression_classifier_iris.onnx", "nusvc_classifier_iris.onnx", "k-nn_classifier_iris.onnx", "linearsvc_classifier_iris.onnx", "adaboost_classifier_iris.onnx", "passive-aggressive_classifier_iris.onnx", "gaussian_naive_bayes_classifier_iris.onnx", "multinomial_naive_bayes_classifier_iris.onnx", "sgd_classifier_iris.onnx", "categorical_naive_bayes_classifier_iris.onnx", "ridge_classifier_iris.onnx", "ridge_classifiercv_iris.onnx", "complement_naive_bayes_classifier_iris.onnx", "perceptron_classifier_iris.onnx", "bernoulli_naive_bayes_classifier_iris.onnx" }; //--- test all iris dataset sample by sample for(int i=0; i<ArraySize(iris_models); i++) { //--- load ONNX-model string model_name="IRIS_models\\"+iris_models[i]; //--- long model=OnnxCreate(model_name,0); if(model==INVALID_HANDLE) { PrintFormat("model_name=%s OnnxCreate error %d for",model_name,GetLastError()); } else { //--- check all samples int correct_results=0; for(int k=0; k<total_samples; k++) { int model_class_id=-1; //--- select data output processor string current_model=iris_models[i]; if(current_model=="svc_classifier_iris.onnx" || current_model=="linearsvc_classifier_iris.onnx" || current_model=="nusvc_classifier_iris.onnx" || current_model=="ridge_classifier_iris.onnx" || current_model=="ridge_classifiercv_iris.onnx" || current_model=="radius_neighbors_classifier_iris.onnx") { TestSampleTensorOutput(model,iris_samples[k],model_class_id); } else { TestSampleSequenceMapOutput(model,iris_samples[k],model_class_id); } //--- if(model_class_id==iris_samples[k].class_id) { correct_results++; //PrintFormat("sample=%d OK [class=%d]",iris_samples[k].sample_id,model_class_id); } else { //PrintFormat("model:%s sample=%d FAILED [class=%d, true class=%d] features=(%.2f,%.2f,%.2f,%.2f]",model_name,iris_samples[k].sample_id,model_class_id,iris_samples[k].class_id,iris_samples[k].features[0],iris_samples[k].features[1],iris_samples[k].features[2],iris_samples[k].features[3]); } } PrintFormat("%d model:%s accuracy: %.4f",i+1,model_name,1.0*correct_results/total_samples); //--- release model OnnxRelease(model); } //--- } return(0); } //+------------------------------------------------------------------+
输出：
Iris_AllClassifiers (EURUSD,H1) 1 model:IRIS_models\random_forest_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 2 model:IRIS_models\gradient_boosting_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 3 model:IRIS_models\bagging_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 4 model:IRIS_models\decision_tree_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 5 model:IRIS_models\extra_tree_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 6 model:IRIS_models\extra_trees_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 7 model:IRIS_models\hist_gradient_boosting_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 8 model:IRIS_models\logistic_regressioncv_classifier_iris.onnx accuracy: 0.9800 Iris_AllClassifiers (EURUSD,H1) 9 model:IRIS_models\mlp_classifier_iris.onnx accuracy: 0.9800 Iris_AllClassifiers (EURUSD,H1) 10 model:IRIS_models\linear_discriminant_analysis_classifier_iris.onnx accuracy: 0.9800 Iris_AllClassifiers (EURUSD,H1) 11 model:IRIS_models\svc_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) 12 model:IRIS_models\radius_neighbors_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) 13 model:IRIS_models\logistic_regression_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) 14 model:IRIS_models\nusvc_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) 15 model:IRIS_models\k-nn_classifier_iris.onnx accuracy: 0.9667 Iris_AllClassifiers (EURUSD,H1) 16 model:IRIS_models\linearsvc_classifier_iris.onnx accuracy: 0.9667 Iris_AllClassifiers (EURUSD,H1) 17 model:IRIS_models\adaboost_classifier_iris.onnx accuracy: 0.9600 Iris_AllClassifiers (EURUSD,H1) 18 model:IRIS_models\passive-aggressive_classifier_iris.onnx accuracy: 0.9600 Iris_AllClassifiers (EURUSD,H1) 19 model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx accuracy: 0.9600 Iris_AllClassifiers (EURUSD,H1) 20 model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx accuracy: 0.9533 Iris_AllClassifiers (EURUSD,H1) 21 model:IRIS_models\sgd_classifier_iris.onnx accuracy: 0.9333 Iris_AllClassifiers (EURUSD,H1) 22 model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx accuracy: 0.9333 Iris_AllClassifiers (EURUSD,H1) 23 model:IRIS_models\ridge_classifier_iris.onnx accuracy: 0.8533 Iris_AllClassifiers (EURUSD,H1) 24 model:IRIS_models\ridge_classifiercv_iris.onnx accuracy: 0.8533 Iris_AllClassifiers (EURUSD,H1) ONNX: Removing initializer 'class_log_prior'. It is not used by any node and should be removed from the model. Iris_AllClassifiers (EURUSD,H1) 25 model:IRIS_models\complement_naive_bayes_classifier_iris.onnx accuracy: 0.6667 Iris_AllClassifiers (EURUSD,H1) 26 model:IRIS_models\perceptron_classifier_iris.onnx accuracy: 0.6133 Iris_AllClassifiers (EURUSD,H1) 27 model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx accuracy: 0.3333
与脚本2.28.1.1的结果对比：
Python Random Forest Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 1 model:IRIS_models\random_forest_classifier_iris.onnx accuracy: 1.0000 Python Gradient Boosting Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 2 model:IRIS_models\gradient_boosting_classifier_iris.onnx accuracy: 1.0000 Python Bagging Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 3 model:IRIS_models\bagging_classifier_iris.onnx accuracy: 1.0000 Python Decision Tree Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 4 model:IRIS_models\decision_tree_classifier_iris.onnx accuracy: 1.0000 Python Extra Tree Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 5 model:IRIS_models\extra_tree_classifier_iris.onnx accuracy: 1.0000 Python Extra Trees Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 6 model:IRIS_models\extra_trees_classifier_iris.onnx accuracy: 1.0000 Python Hist Gradient Boosting Classifier - Original Accuracy: 1.0000, ONNX Accuracy: 1.0000 7 model:IRIS_models\hist_gradient_boosting_classifier_iris.onnx accuracy: 1.0000 Python Logistic RegressionCV Classifier - Original Accuracy: 0.9800, ONNX Accuracy: 0.9800 8 model:IRIS_models\logistic_regressioncv_classifier_iris.onnx accuracy: 0.9800 Python MLP Classifier - Original Accuracy: 0.9800, ONNX Accuracy: 0.9800 9 model:IRIS_models\mlp_classifier_iris.onnx accuracy: 0.9800 Python Linear Discriminant Analysis Classifier - Original Accuracy: 0.9800, ONNX Accuracy: 0.9800 10 model:IRIS_models\linear_discriminant_analysis_classifier_iris.onnx accuracy: 0.9800 Python SVC Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 11 model:IRIS_models\svc_classifier_iris.onnx accuracy: 0.9733 Python Radius Neighbors Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 12 model:IRIS_models\radius_neighbors_classifier_iris.onnx accuracy: 0.9733 Python Logistic Regression Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 13 model:IRIS_models\logistic_regression_classifier_iris.onnx accuracy: 0.9733 Python NuSVC Classifier - Original Accuracy: 0.9733, ONNX Accuracy: 0.9733 14 model:IRIS_models\nusvc_classifier_iris.onnx accuracy: 0.9733 Python K-NN Classifier - Original Accuracy: 0.9667, ONNX Accuracy: 0.9667 15 model:IRIS_models\k-nn_classifier_iris.onnx accuracy: 0.9667 Python LinearSVC Classifier - Original Accuracy: 0.9667, ONNX Accuracy: 0.9667 16 model:IRIS_models\linearsvc_classifier_iris.onnx accuracy: 0.9667 Python AdaBoost Classifier - Original Accuracy: 0.9600, ONNX Accuracy: 0.9600 17 model:IRIS_models\adaboost_classifier_iris.onnx accuracy: 0.9600 Python Passive-Aggressive Classifier - Original Accuracy: 0.9600, ONNX Accuracy: 0.9600 18 model:IRIS_models\passive-aggressive_classifier_iris.onnx accuracy: 0.9600 Python Gaussian Naive Bayes Classifier - Original Accuracy: 0.9600, ONNX Accuracy: 0.9600 19 model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx accuracy: 0.9600 Python Multinomial Naive Bayes Classifier - Original Accuracy: 0.9533, ONNX Accuracy: 0.9533 20 model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx accuracy: 0.9533 Python SGD Classifier - Original Accuracy: 0.9333, ONNX Accuracy: 0.9333 21 model:IRIS_models\sgd_classifier_iris.onnx accuracy: 0.9333 Python Categorical Naive Bayes Classifier - Original Accuracy: 0.9333, ONNX Accuracy: 0.9333 22 model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx accuracy: 0.9333 Python Ridge Classifier - Original Accuracy: 0.8533, ONNX Accuracy: 0.8533 23 model:IRIS_models\ridge_classifier_iris.onnx accuracy: 0.8533 Python Ridge ClassifierCV - Original Accuracy: 0.8533, ONNX Accuracy: 0.8533 24 model:IRIS_models\ridge_classifiercv_iris.onnx accuracy: 0.8533 Python Complement Naive Bayes Classifier - Original Accuracy: 0.6667, ONNX Accuracy: 0.6667 25 model:IRIS_models\complement_naive_bayes_classifier_iris.onnx accuracy: 0.6667 Python Perceptron Classifier - Original Accuracy: 0.6133, ONNX Accuracy: 0.6133 26 model:IRIS_models\perceptron_classifier_iris.onnx accuracy: 0.6133 Python Bernoulli Naive Bayes Classifier - Original Accuracy: 0.3333, ONNX Accuracy: 0.3333 27 model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx accuracy: 0.3333
值得注意的是，在 MQL5 中执行所有保存的 ONNX 模型与 2.28.1的结果完全对应。
因此，我们检查的模型在转换为 ONNX 格式后保留了其分类准确性。
值得一提的是，七个模型对鸢尾花数据集实现了完美的分类准确率（准确率=1.0）：
- Random Forest Classifier;
- Gradient Boosting Classifier;
- Bagging Classifier;
- Decision Tree Classifier;
- Extra Tree Classifier;
- Extra Trees Classifier;
- Histogram Gradient Boosting Classifier.
其余 20 个模型出现分类错误。
如果取消注释第 208 行，脚本还将显示每个模型错误分类的鸢尾花数据集样本：
Iris_AllClassifiers (EURUSD,H1) 1 model:IRIS_models\random_forest_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 2 model:IRIS_models\gradient_boosting_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 3 model:IRIS_models\bagging_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 4 model:IRIS_models\decision_tree_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 5 model:IRIS_models\extra_tree_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 6 model:IRIS_models\extra_trees_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) 7 model:IRIS_models\hist_gradient_boosting_classifier_iris.onnx accuracy: 1.0000 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regressioncv_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regressioncv_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regressioncv_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) 8 model:IRIS_models\logistic_regressioncv_classifier_iris.onnx accuracy: 0.9800 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\mlp_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\mlp_classifier_iris.onnx sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\mlp_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) 9 model:IRIS_models\mlp_classifier_iris.onnx accuracy: 0.9800 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linear_discriminant_analysis_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linear_discriminant_analysis_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linear_discriminant_analysis_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) 10 model:IRIS_models\linear_discriminant_analysis_classifier_iris.onnx accuracy: 0.9800 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\svc_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\svc_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\svc_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\svc_classifier_iris.onnx sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) 11 model:IRIS_models\svc_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\radius_neighbors_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\radius_neighbors_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\radius_neighbors_classifier_iris.onnx sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\radius_neighbors_classifier_iris.onnx sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) 12 model:IRIS_models\radius_neighbors_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regression_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regression_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regression_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\logistic_regression_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) 13 model:IRIS_models\logistic_regression_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\nusvc_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\nusvc_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\nusvc_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\nusvc_classifier_iris.onnx sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) 14 model:IRIS_models\nusvc_classifier_iris.onnx accuracy: 0.9733 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\k-nn_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\k-nn_classifier_iris.onnx sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\k-nn_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\k-nn_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\k-nn_classifier_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) 15 model:IRIS_models\k-nn_classifier_iris.onnx accuracy: 0.9667 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linearsvc_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linearsvc_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linearsvc_classifier_iris.onnx sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linearsvc_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\linearsvc_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) 16 model:IRIS_models\linearsvc_classifier_iris.onnx accuracy: 0.9667 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\adaboost_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\adaboost_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\adaboost_classifier_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\adaboost_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\adaboost_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\adaboost_classifier_iris.onnx sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AllClassifiers (EURUSD,H1) 17 model:IRIS_models\adaboost_classifier_iris.onnx accuracy: 0.9600 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\passive-aggressive_classifier_iris.onnx sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\passive-aggressive_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\passive-aggressive_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\passive-aggressive_classifier_iris.onnx sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\passive-aggressive_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\passive-aggressive_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) 18 model:IRIS_models\passive-aggressive_classifier_iris.onnx accuracy: 0.9600 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) 19 model:IRIS_models\gaussian_naive_bayes_classifier_iris.onnx accuracy: 0.9600 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=69 FAILED [class=2, true class=1] features=(6.20,2.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=132 FAILED [class=1, true class=2] features=(7.90,3.80,6.40,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) 20 model:IRIS_models\multinomial_naive_bayes_classifier_iris.onnx accuracy: 0.9533 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=65 FAILED [class=0, true class=1] features=(5.60,2.90,3.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=86 FAILED [class=0, true class=1] features=(6.00,3.40,4.50,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=124 FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\sgd_classifier_iris.onnx sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AllClassifiers (EURUSD,H1) 21 model:IRIS_models\sgd_classifier_iris.onnx accuracy: 0.9333 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=102 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=122 FAILED [class=1, true class=2] features=(5.60,2.80,4.90,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=124 FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=128 FAILED [class=1, true class=2] features=(6.10,3.00,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx sample=143 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_AllClassifiers (EURUSD,H1) 22 model:IRIS_models\categorical_naive_bayes_classifier_iris.onnx accuracy: 0.9333 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=51 FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=52 FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=57 FAILED [class=2, true class=1] features=(6.30,3.30,4.70,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=62 FAILED [class=2, true class=1] features=(5.90,3.00,4.20,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=65 FAILED [class=2, true class=1] features=(5.60,2.90,3.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=66 FAILED [class=2, true class=1] features=(6.70,3.10,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=76 FAILED [class=2, true class=1] features=(6.60,3.00,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=79 FAILED [class=2, true class=1] features=(6.00,2.90,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=86 FAILED [class=2, true class=1] features=(6.00,3.40,4.50,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=87 FAILED [class=2, true class=1] features=(6.70,3.10,4.70,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=89 FAILED [class=2, true class=1] features=(5.60,3.00,4.10,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=92 FAILED [class=2, true class=1] features=(6.10,3.00,4.60,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=109 FAILED [class=1, true class=2] features=(6.70,2.50,5.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifier_iris.onnx sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AllClassifiers (EURUSD,H1) 23 model:IRIS_models\ridge_classifier_iris.onnx accuracy: 0.8533 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=51 FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=52 FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=57 FAILED [class=2, true class=1] features=(6.30,3.30,4.70,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=62 FAILED [class=2, true class=1] features=(5.90,3.00,4.20,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=65 FAILED [class=2, true class=1] features=(5.60,2.90,3.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=66 FAILED [class=2, true class=1] features=(6.70,3.10,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=76 FAILED [class=2, true class=1] features=(6.60,3.00,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=79 FAILED [class=2, true class=1] features=(6.00,2.90,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=86 FAILED [class=2, true class=1] features=(6.00,3.40,4.50,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=87 FAILED [class=2, true class=1] features=(6.70,3.10,4.70,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=89 FAILED [class=2, true class=1] features=(5.60,3.00,4.10,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=92 FAILED [class=2, true class=1] features=(6.10,3.00,4.60,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=109 FAILED [class=1, true class=2] features=(6.70,2.50,5.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\ridge_classifiercv_iris.onnx sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AllClassifiers (EURUSD,H1) 24 model:IRIS_models\ridge_classifiercv_iris.onnx accuracy: 0.8533 Iris_AllClassifiers (EURUSD,H1) ONNX: Removing initializer 'class_log_prior'. It is not used by any node and should be removed from the model. Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=51 FAILED [class=2, true class=1] features=(7.00,3.20,4.70,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=52 FAILED [class=2, true class=1] features=(6.40,3.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=53 FAILED [class=2, true class=1] features=(6.90,3.10,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=54 FAILED [class=2, true class=1] features=(5.50,2.30,4.00,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=55 FAILED [class=2, true class=1] features=(6.50,2.80,4.60,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=56 FAILED [class=2, true class=1] features=(5.70,2.80,4.50,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=57 FAILED [class=2, true class=1] features=(6.30,3.30,4.70,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=58 FAILED [class=2, true class=1] features=(4.90,2.40,3.30,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=59 FAILED [class=2, true class=1] features=(6.60,2.90,4.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=60 FAILED [class=2, true class=1] features=(5.20,2.70,3.90,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=61 FAILED [class=2, true class=1] features=(5.00,2.00,3.50,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=62 FAILED [class=2, true class=1] features=(5.90,3.00,4.20,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=63 FAILED [class=2, true class=1] features=(6.00,2.20,4.00,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=64 FAILED [class=2, true class=1] features=(6.10,2.90,4.70,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=65 FAILED [class=2, true class=1] features=(5.60,2.90,3.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=66 FAILED [class=2, true class=1] features=(6.70,3.10,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=67 FAILED [class=2, true class=1] features=(5.60,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=68 FAILED [class=2, true class=1] features=(5.80,2.70,4.10,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=69 FAILED [class=2, true class=1] features=(6.20,2.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=70 FAILED [class=2, true class=1] features=(5.60,2.50,3.90,1.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=71 FAILED [class=2, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=72 FAILED [class=2, true class=1] features=(6.10,2.80,4.00,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=73 FAILED [class=2, true class=1] features=(6.30,2.50,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=74 FAILED [class=2, true class=1] features=(6.10,2.80,4.70,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=75 FAILED [class=2, true class=1] features=(6.40,2.90,4.30,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=76 FAILED [class=2, true class=1] features=(6.60,3.00,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=77 FAILED [class=2, true class=1] features=(6.80,2.80,4.80,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=78 FAILED [class=2, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=79 FAILED [class=2, true class=1] features=(6.00,2.90,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=80 FAILED [class=0, true class=1] features=(5.70,2.60,3.50,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=81 FAILED [class=2, true class=1] features=(5.50,2.40,3.80,1.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=82 FAILED [class=2, true class=1] features=(5.50,2.40,3.70,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=83 FAILED [class=2, true class=1] features=(5.80,2.70,3.90,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=84 FAILED [class=2, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=85 FAILED [class=2, true class=1] features=(5.40,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=86 FAILED [class=2, true class=1] features=(6.00,3.40,4.50,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=87 FAILED [class=2, true class=1] features=(6.70,3.10,4.70,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=88 FAILED [class=2, true class=1] features=(6.30,2.30,4.40,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=89 FAILED [class=2, true class=1] features=(5.60,3.00,4.10,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=90 FAILED [class=2, true class=1] features=(5.50,2.50,4.00,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=91 FAILED [class=2, true class=1] features=(5.50,2.60,4.40,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=92 FAILED [class=2, true class=1] features=(6.10,3.00,4.60,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=93 FAILED [class=2, true class=1] features=(5.80,2.60,4.00,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=94 FAILED [class=2, true class=1] features=(5.00,2.30,3.30,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=95 FAILED [class=2, true class=1] features=(5.60,2.70,4.20,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=96 FAILED [class=2, true class=1] features=(5.70,3.00,4.20,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=97 FAILED [class=2, true class=1] features=(5.70,2.90,4.20,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=98 FAILED [class=2, true class=1] features=(6.20,2.90,4.30,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=99 FAILED [class=0, true class=1] features=(5.10,2.50,3.00,1.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\complement_naive_bayes_classifier_iris.onnx sample=100 FAILED [class=2, true class=1] features=(5.70,2.80,4.10,1.30] Iris_AllClassifiers (EURUSD,H1) 25 model:IRIS_models\complement_naive_bayes_classifier_iris.onnx accuracy: 0.6667 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=2 FAILED [class=1, true class=0] features=(4.90,3.00,1.40,0.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=9 FAILED [class=1, true class=0] features=(4.40,2.90,1.40,0.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=10 FAILED [class=1, true class=0] features=(4.90,3.10,1.50,0.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=13 FAILED [class=1, true class=0] features=(4.80,3.00,1.40,0.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=21 FAILED [class=1, true class=0] features=(5.40,3.40,1.70,0.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=26 FAILED [class=1, true class=0] features=(5.00,3.00,1.60,0.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=31 FAILED [class=1, true class=0] features=(4.80,3.10,1.60,0.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=35 FAILED [class=1, true class=0] features=(4.90,3.10,1.50,0.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=42 FAILED [class=1, true class=0] features=(4.50,2.30,1.30,0.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=46 FAILED [class=1, true class=0] features=(4.80,3.00,1.40,0.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=102 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=103 FAILED [class=1, true class=2] features=(7.10,3.00,5.90,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=104 FAILED [class=1, true class=2] features=(6.30,2.90,5.60,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=105 FAILED [class=1, true class=2] features=(6.50,3.00,5.80,2.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=106 FAILED [class=1, true class=2] features=(7.60,3.00,6.60,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=107 FAILED [class=1, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=108 FAILED [class=1, true class=2] features=(7.30,2.90,6.30,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=109 FAILED [class=1, true class=2] features=(6.70,2.50,5.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=110 FAILED [class=1, true class=2] features=(7.20,3.60,6.10,2.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=111 FAILED [class=1, true class=2] features=(6.50,3.20,5.10,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=112 FAILED [class=1, true class=2] features=(6.40,2.70,5.30,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=113 FAILED [class=1, true class=2] features=(6.80,3.00,5.50,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=114 FAILED [class=1, true class=2] features=(5.70,2.50,5.00,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=116 FAILED [class=1, true class=2] features=(6.40,3.20,5.30,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=117 FAILED [class=1, true class=2] features=(6.50,3.00,5.50,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=118 FAILED [class=1, true class=2] features=(7.70,3.80,6.70,2.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=119 FAILED [class=1, true class=2] features=(7.70,2.60,6.90,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=120 FAILED [class=1, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=121 FAILED [class=1, true class=2] features=(6.90,3.20,5.70,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=122 FAILED [class=1, true class=2] features=(5.60,2.80,4.90,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=123 FAILED [class=1, true class=2] features=(7.70,2.80,6.70,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=124 FAILED [class=1, true class=2] features=(6.30,2.70,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=125 FAILED [class=1, true class=2] features=(6.70,3.30,5.70,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=126 FAILED [class=1, true class=2] features=(7.20,3.20,6.00,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=127 FAILED [class=1, true class=2] features=(6.20,2.80,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=128 FAILED [class=1, true class=2] features=(6.10,3.00,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=129 FAILED [class=1, true class=2] features=(6.40,2.80,5.60,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=130 FAILED [class=1, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=131 FAILED [class=1, true class=2] features=(7.40,2.80,6.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=132 FAILED [class=1, true class=2] features=(7.90,3.80,6.40,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=133 FAILED [class=1, true class=2] features=(6.40,2.80,5.60,2.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=134 FAILED [class=1, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=135 FAILED [class=1, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=136 FAILED [class=1, true class=2] features=(7.70,3.00,6.10,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=137 FAILED [class=1, true class=2] features=(6.30,3.40,5.60,2.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=138 FAILED [class=1, true class=2] features=(6.40,3.10,5.50,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=139 FAILED [class=1, true class=2] features=(6.00,3.00,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=140 FAILED [class=1, true class=2] features=(6.90,3.10,5.40,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=141 FAILED [class=1, true class=2] features=(6.70,3.10,5.60,2.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=142 FAILED [class=1, true class=2] features=(6.90,3.10,5.10,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=143 FAILED [class=1, true class=2] features=(5.80,2.70,5.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=144 FAILED [class=1, true class=2] features=(6.80,3.20,5.90,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=145 FAILED [class=1, true class=2] features=(6.70,3.30,5.70,2.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=146 FAILED [class=1, true class=2] features=(6.70,3.00,5.20,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=147 FAILED [class=1, true class=2] features=(6.30,2.50,5.00,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=148 FAILED [class=1, true class=2] features=(6.50,3.00,5.20,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=149 FAILED [class=1, true class=2] features=(6.20,3.40,5.40,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\perceptron_classifier_iris.onnx sample=150 FAILED [class=1, true class=2] features=(5.90,3.00,5.10,1.80] Iris_AllClassifiers (EURUSD,H1) 26 model:IRIS_models\perceptron_classifier_iris.onnx accuracy: 0.6133 Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=51 FAILED [class=0, true class=1] features=(7.00,3.20,4.70,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=52 FAILED [class=0, true class=1] features=(6.40,3.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=53 FAILED [class=0, true class=1] features=(6.90,3.10,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=54 FAILED [class=0, true class=1] features=(5.50,2.30,4.00,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=55 FAILED [class=0, true class=1] features=(6.50,2.80,4.60,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=56 FAILED [class=0, true class=1] features=(5.70,2.80,4.50,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=57 FAILED [class=0, true class=1] features=(6.30,3.30,4.70,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=58 FAILED [class=0, true class=1] features=(4.90,2.40,3.30,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=59 FAILED [class=0, true class=1] features=(6.60,2.90,4.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=60 FAILED [class=0, true class=1] features=(5.20,2.70,3.90,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=61 FAILED [class=0, true class=1] features=(5.00,2.00,3.50,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=62 FAILED [class=0, true class=1] features=(5.90,3.00,4.20,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=63 FAILED [class=0, true class=1] features=(6.00,2.20,4.00,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=64 FAILED [class=0, true class=1] features=(6.10,2.90,4.70,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=65 FAILED [class=0, true class=1] features=(5.60,2.90,3.60,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=66 FAILED [class=0, true class=1] features=(6.70,3.10,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=67 FAILED [class=0, true class=1] features=(5.60,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=68 FAILED [class=0, true class=1] features=(5.80,2.70,4.10,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=69 FAILED [class=0, true class=1] features=(6.20,2.20,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=70 FAILED [class=0, true class=1] features=(5.60,2.50,3.90,1.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=71 FAILED [class=0, true class=1] features=(5.90,3.20,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=72 FAILED [class=0, true class=1] features=(6.10,2.80,4.00,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=73 FAILED [class=0, true class=1] features=(6.30,2.50,4.90,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=74 FAILED [class=0, true class=1] features=(6.10,2.80,4.70,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=75 FAILED [class=0, true class=1] features=(6.40,2.90,4.30,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=76 FAILED [class=0, true class=1] features=(6.60,3.00,4.40,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=77 FAILED [class=0, true class=1] features=(6.80,2.80,4.80,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=78 FAILED [class=0, true class=1] features=(6.70,3.00,5.00,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=79 FAILED [class=0, true class=1] features=(6.00,2.90,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=80 FAILED [class=0, true class=1] features=(5.70,2.60,3.50,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=81 FAILED [class=0, true class=1] features=(5.50,2.40,3.80,1.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=82 FAILED [class=0, true class=1] features=(5.50,2.40,3.70,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=83 FAILED [class=0, true class=1] features=(5.80,2.70,3.90,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=84 FAILED [class=0, true class=1] features=(6.00,2.70,5.10,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=85 FAILED [class=0, true class=1] features=(5.40,3.00,4.50,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=86 FAILED [class=0, true class=1] features=(6.00,3.40,4.50,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=87 FAILED [class=0, true class=1] features=(6.70,3.10,4.70,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=88 FAILED [class=0, true class=1] features=(6.30,2.30,4.40,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=89 FAILED [class=0, true class=1] features=(5.60,3.00,4.10,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=90 FAILED [class=0, true class=1] features=(5.50,2.50,4.00,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=91 FAILED [class=0, true class=1] features=(5.50,2.60,4.40,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=92 FAILED [class=0, true class=1] features=(6.10,3.00,4.60,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=93 FAILED [class=0, true class=1] features=(5.80,2.60,4.00,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=94 FAILED [class=0, true class=1] features=(5.00,2.30,3.30,1.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=95 FAILED [class=0, true class=1] features=(5.60,2.70,4.20,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=96 FAILED [class=0, true class=1] features=(5.70,3.00,4.20,1.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=97 FAILED [class=0, true class=1] features=(5.70,2.90,4.20,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=98 FAILED [class=0, true class=1] features=(6.20,2.90,4.30,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=99 FAILED [class=0, true class=1] features=(5.10,2.50,3.00,1.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=100 FAILED [class=0, true class=1] features=(5.70,2.80,4.10,1.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=101 FAILED [class=0, true class=2] features=(6.30,3.30,6.00,2.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=102 FAILED [class=0, true class=2] features=(5.80,2.70,5.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=103 FAILED [class=0, true class=2] features=(7.10,3.00,5.90,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=104 FAILED [class=0, true class=2] features=(6.30,2.90,5.60,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=105 FAILED [class=0, true class=2] features=(6.50,3.00,5.80,2.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=106 FAILED [class=0, true class=2] features=(7.60,3.00,6.60,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=107 FAILED [class=0, true class=2] features=(4.90,2.50,4.50,1.70] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=108 FAILED [class=0, true class=2] features=(7.30,2.90,6.30,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=109 FAILED [class=0, true class=2] features=(6.70,2.50,5.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=110 FAILED [class=0, true class=2] features=(7.20,3.60,6.10,2.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=111 FAILED [class=0, true class=2] features=(6.50,3.20,5.10,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=112 FAILED [class=0, true class=2] features=(6.40,2.70,5.30,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=113 FAILED [class=0, true class=2] features=(6.80,3.00,5.50,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=114 FAILED [class=0, true class=2] features=(5.70,2.50,5.00,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=115 FAILED [class=0, true class=2] features=(5.80,2.80,5.10,2.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=116 FAILED [class=0, true class=2] features=(6.40,3.20,5.30,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=117 FAILED [class=0, true class=2] features=(6.50,3.00,5.50,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=118 FAILED [class=0, true class=2] features=(7.70,3.80,6.70,2.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=119 FAILED [class=0, true class=2] features=(7.70,2.60,6.90,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=120 FAILED [class=0, true class=2] features=(6.00,2.20,5.00,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=121 FAILED [class=0, true class=2] features=(6.90,3.20,5.70,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=122 FAILED [class=0, true class=2] features=(5.60,2.80,4.90,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=123 FAILED [class=0, true class=2] features=(7.70,2.80,6.70,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=124 FAILED [class=0, true class=2] features=(6.30,2.70,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=125 FAILED [class=0, true class=2] features=(6.70,3.30,5.70,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=126 FAILED [class=0, true class=2] features=(7.20,3.20,6.00,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=127 FAILED [class=0, true class=2] features=(6.20,2.80,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=128 FAILED [class=0, true class=2] features=(6.10,3.00,4.90,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=129 FAILED [class=0, true class=2] features=(6.40,2.80,5.60,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=130 FAILED [class=0, true class=2] features=(7.20,3.00,5.80,1.60] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=131 FAILED [class=0, true class=2] features=(7.40,2.80,6.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=132 FAILED [class=0, true class=2] features=(7.90,3.80,6.40,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=133 FAILED [class=0, true class=2] features=(6.40,2.80,5.60,2.20] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=134 FAILED [class=0, true class=2] features=(6.30,2.80,5.10,1.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=135 FAILED [class=0, true class=2] features=(6.10,2.60,5.60,1.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=136 FAILED [class=0, true class=2] features=(7.70,3.00,6.10,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=137 FAILED [class=0, true class=2] features=(6.30,3.40,5.60,2.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=138 FAILED [class=0, true class=2] features=(6.40,3.10,5.50,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=139 FAILED [class=0, true class=2] features=(6.00,3.00,4.80,1.80] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=140 FAILED [class=0, true class=2] features=(6.90,3.10,5.40,2.10] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=141 FAILED [class=0, true class=2] features=(6.70,3.10,5.60,2.40] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=142 FAILED [class=0, true class=2] features=(6.90,3.10,5.10,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=143 FAILED [class=0, true class=2] features=(5.80,2.70,5.10,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=144 FAILED [class=0, true class=2] features=(6.80,3.20,5.90,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=145 FAILED [class=0, true class=2] features=(6.70,3.30,5.70,2.50] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=146 FAILED [class=0, true class=2] features=(6.70,3.00,5.20,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=147 FAILED [class=0, true class=2] features=(6.30,2.50,5.00,1.90] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=148 FAILED [class=0, true class=2] features=(6.50,3.00,5.20,2.00] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=149 FAILED [class=0, true class=2] features=(6.20,3.40,5.40,2.30] Iris_AllClassifiers (EURUSD,H1) model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx sample=150 FAILED [class=0, true class=2] features=(5.90,3.00,5.10,1.80] Iris_AllClassifiers (EURUSD,H1) 27 model:IRIS_models\bernoulli_naive_bayes_classifier_iris.onnx accuracy: 0.3333
2.29.Scikit-Learn 中无法转换为 ONNX 的分类模型
由于 convert_sklearn 过程中的错误，某些分类模型无法转换为 ONNX 格式。
2.29.1.DummyClassifier
DummyClassifier 是 Scikit-learn 库中的一个分类器，用作分类任务的简单基准模型。它旨在测试和评估更复杂的分类模型的性能。
工作准则：
DummyClassifier 的工作原理非常简单；它不考虑输入数据而做出随机或简单的预测。它提供不同的策略（通过“strategy”参数选择）：
- “most_frequent”（最常见的类别）：该策略总是预测训练数据集中出现频率最高的类别。在类别不平衡且您需要预测主导类别的情况下，它会很有用。
- “stratified”(分层选择)：该策略尝试做出与训练数据集中的类别分布相匹配的预测。它使用随机猜测，但考虑到类别比例。
- “uniform”（均匀分布）：该策略对每个类别以相同的概率做出随机预测。当类别平衡并且您想要测试模型的平均表现时，它很有用。
能力：
- 简单性：DummyClassifier 可用于测试训练基准模型的速度以及它产生的结果。它有助于快速评估其他分类器的性能。
- 流水线用途：您可以使用 DummyClassifier 作为流水线中的基准模型，并结合其他转换和模型进行比较和测试。
限制：
- 不使用数据：DummyClassifier 在没有考虑实际数据的情况下做出随机或简单的预测。它无法从数据中学习或发现模式。
- 不适合复杂任务：该分类器不是为解决复杂的分类任务而设计的，并且通常对于具有大型数据集和复杂模式的任务不会产生良好的结果。
- 缺乏信息量：使用 DummyClassifier 获得的结果可能缺乏信息量，并且无法提供有关模型性能的有用信息。它们对于代码测试和评估更有用。
DummyClassifier 是初始测试和评估分类模型的有用工具，但其在复杂任务中的用途有限，并且不能取代更高级的分类算法。
2.29.1.1.创建 DummyClassifier 模型的代码
# Iris_DummyClassifier.py
# The code demonstrates the process of training DummyClassifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model.
# It also evaluates the accuracy of both the original model and the ONNX model.
# Copyright 2023, MetaQuotes Ltd.
# https://www.mql5.com
# import necessary libraries
from sklearn import datasets
from sklearn.dummy import DummyClassifier
from sklearn.metrics import accuracy_score, classification_report
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as ort
import numpy as np
from sys import argv
# define the path for saving the model
data_path = argv[0]
last_index = data_path.rfind("\\") + 1
data_path = data_path[0:last_index]
# load the Iris dataset
iris = datasets.load_iris()
X = iris.data
y = iris.target
# create a DummyClassifier model with the strategy "most_frequent"
dummy_classifier = DummyClassifier(strategy="most_frequent")
# train the model on the entire dataset
dummy_classifier.fit(X, y)
# predict classes for the entire dataset
y_pred = dummy_classifier.predict(X)
# evaluate the model's accuracy
accuracy = accuracy_score(y, y_pred)
print("Accuracy of DummyClassifier model:", accuracy)
# display the classification report
print("\nClassification Report:\n", classification_report(y, y_pred))
# define the input data type
initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))]
# export the model to ONNX format with float data type
onnx_model = convert_sklearn(dummy_classifier, initial_types=initial_type, target_opset=12)
# save the model to a file
onnx_filename = data_path + "dummy_classifier_iris.onnx"
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
# print model path
print(f"Model saved to {onnx_filename}")
# load the ONNX model and make predictions
onnx_session = ort.InferenceSession(onnx_filename)
input_name = onnx_session.get_inputs()[0].name
output_name = onnx_session.get_outputs()[0].name
# display information about input tensors in ONNX
print("\nInformation about input tensors in ONNX:")
for i, input_tensor in enumerate(onnx_session.get_inputs()):
print(f"{i + 1}.Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}")
# display information about output tensors in ONNX
print("\nInformation about output tensors in ONNX:")
for i, output_tensor in enumerate(onnx_session.get_outputs()):
print(f"{i + 1}.Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}")
# convert data to floating-point format (float32)
X_float32 = X.astype(np.float32)
# predict classes for the entire dataset using ONNX
y_pred_onnx = onnx_session.run([output_name], {input_name:X_float32})[0]
# evaluate the accuracy of the ONNX model
accuracy_onnx = accuracy_score(y, y_pred_onnx)
print("\nAccuracy of DummyClassifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of DummyClassifier model:0.3333333333333333
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.33 1.00 0.50 50
Python 1 0.00 0.00 0.00 50
Python 2 0.00 0.00 0.00 50
Python
Python accuracy 0.33 150
Python macro avg 0.11 0.33 0.17 150
Python weighted avg 0.11 0.33 0.17 150
Python
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 0.33 1.00 0.50 50
Python 1 0.00 0.00 0.00 50
Python 2 0.00 0.00 0.00 50
Python
Python accuracy 0.33 150
Python macro avg 0.11 0.33 0.17 150
Python weighted avg 0.11 0.33 0.17 150
Python
该模型在 Scikit-learn 中构建并成功运行，但在转换为 ONNX 时出现错误。
在“错误”选项卡中，显示有关转换为 ONNX 格式的错误消息：
onnx_model = convert_sklearn(dummy_classifier, initial_types=initial_type, target_opset=12) Iris_DummyClassifier.py 44 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.dummy.DummyClassifier'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_DummyClassifier.py finished in 2071 ms 19 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.dummy.DummyClassifier'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_DummyClassifier.py finished in 2071 ms 19 1
因此，DummyClassifier 模型无法转换为 ONNX。
2.29.2.GaussianProcessClassifier
GaussianProcessClassifier 是一个使用高斯过程进行分类任务的分类器。它属于使用高斯过程的模型系列，在需要概率类估计的任务中很有用。
工作准则：
- GaussianProcessClassifier 使用高斯过程来模拟从特征空间到类概率估计空间的映射。
- 它通过评估某个点属于每个类的概率来为每个类建立一个概率模型。
- 在分类过程中，它会为给定点选择概率最高的类别。
能力：
- 概率分类：GaussianProcessClassifier 提供概率类估计，这对于评估模型不确定性很有用。
- 适应性：该分类器可以适应数据并根据新的观察结果更新其预测。
- 校准：可以使用“校准”方法校准模型，以改善概率估计。
限制：
- 计算复杂性：对于大型数据集和/或高维特征空间，GaussianProcessClassifier 的计算成本可能很高。
- 不适合大样本：由于计算复杂性，该分类器可能不适用于对大型数据集进行有效的训练。
- 解释复杂性：高斯过程的解释和理解具有挑战性，特别是对于没有贝叶斯统计经验的用户来说。
GaussianProcessClassifier 在概率类估计很重要并且可以处理计算成本的任务中很有价值。否则，对于大数据集或简单数据结构的分类任务，其他分类算法可能更合适。
2.29.2.1. 创建 GaussianProcessClassifier 模型的代码# Iris_GaussianProcessClassifier.py
# The code demonstrates the process of training Iris_GaussianProcess Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model.
# It also evaluates the accuracy of both the original model and the ONNX model.
# Copyright 2023, MetaQuotes Ltd.
# https://www.mql5.com
# import necessary libraries
from sklearn import datasets
from sklearn.gaussian_process import GaussianProcessClassifier
from sklearn.gaussian_process.kernels import RBF
from sklearn.metrics import accuracy_score, classification_report
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as ort
import numpy as np
from sys import argv
# define the path for saving the model
data_path = argv[0]
last_index = data_path.rfind("\\") + 1
data_path = data_path[0:last_index]
# load the Iris dataset
iris = datasets.load_iris()
X = iris.data
y = iris.target
# create a GaussianProcessClassifier model with an RBF kernel
kernel = 1.0 * RBF(1.0)
gpc_model = GaussianProcessClassifier(kernel=kernel)
# train the model on the entire dataset
gpc_model.fit(X, y)
# predict classes for the entire dataset
y_pred = gpc_model.predict(X)
# evaluate the model's accuracy
accuracy = accuracy_score(y, y_pred)
print("Accuracy of GaussianProcessClassifier model:", accuracy)
# display the classification report
print("\nClassification Report:\n", classification_report(y, y_pred))
# define the input data type
initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))]
# export the model to ONNX format with float data type
onnx_model = convert_sklearn(gpc_model, initial_types=initial_type, target_opset=12)
# save the model to a file
onnx_filename = data_path + "gpc_iris.onnx"
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
# print the path to the model
print(f"Model saved to {onnx_filename}")
# load the ONNX model and make predictions
onnx_session = ort.InferenceSession(onnx_filename)
input_name = onnx_session.get_inputs()[0].name
output_name = onnx_session.get_outputs()[0].name
# display information about input tensors in ONNX
print("\nInformation about input tensors in ONNX:")
for i, input_tensor in enumerate(onnx_session.get_inputs()):
print(f"{i + 1}.Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}")
# display information about output tensors in ONNX
print("\nInformation about output tensors in ONNX:")
for i, output_tensor in enumerate(onnx_session.get_outputs()):
print(f"{i + 1}.Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}")
# convert data to floating-point format (float32)
X_float32 = X.astype(np.float32)
# predict classes for the entire dataset using ONNX
y_pred_onnx = onnx_session.run([output_name], {input_name:X_float32})[0]
# evaluate the accuracy of the ONNX model
accuracy_onnx = accuracy_score(y, y_pred_onnx)
print("\nAccuracy of GaussianProcessClassifier model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of GaussianProcessClassifier model:0.9866666666666667
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.98 0.98 50
Python 2 0.98 0.98 0.98 50
Python
Python accuracy 0.99 150
Python macro avg 0.99 0.99 0.99 150
Python weighted avg 0.99 0.99 0.99 150
Python
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.98 0.98 50
Python 2 0.98 0.98 0.98 50
Python
Python accuracy 0.99 150
Python macro avg 0.99 0.99 0.99 150
Python weighted avg 0.99 0.99 0.99 150
Python
在“错误”选项卡中，显示有关转换为 ONNX 格式的错误消息：
onnx_model = convert_sklearn(gpc_model, initial_types=initial_type, target_opset=12) Iris_GaussianProcessClassifier.py 46 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_converter(operator, container, verbose=verbose) _topology.py 1349 1
conv(self.scopes[0], operator, container) _topology.py 1132 1
return self._fct(*args) _registration.py 27 1
raise NotImplementedError("Only binary classification is iplemented.") gaussian_process.py 247 1
NotImplementedError:Only binary classification is iplemented. gaussian_process.py 247 1
Iris_GaussianProcessClassifier.py finished in 4004 ms 9 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_converter(operator, container, verbose=verbose) _topology.py 1349 1
conv(self.scopes[0], operator, container) _topology.py 1132 1
return self._fct(*args) _registration.py 27 1
raise NotImplementedError("Only binary classification is iplemented.") gaussian_process.py 247 1
NotImplementedError:Only binary classification is iplemented. gaussian_process.py 247 1
Iris_GaussianProcessClassifier.py finished in 4004 ms 9 1
因此，GaussianProcessClassifier 模型无法转换为 ONNX。
2.29.3.LabelPropagation Classifier
LabelPropagation 是一种用于分类任务的半监督学习方法。该方法背后的主要思想是在基于图形的数据结构中将标签（类）从标记实例传播到未标记实例。
标签传播过程：
- 它首先构建一个图，其中节点代表数据实例，节点之间的边反映实例之间的相似性或接近性。
- 初始标签分配：标记的实例被赋予了其标签，而未标记的实例则以一些未定义的标签开始。
- 图上的标签传播：根据实例之间的相似性，标记实例的标签会传播到未标记实例。这种相似性可以通过多种方式确定，例如，使用图中的最近邻居。
- 迭代过程：标签可以在几次迭代中发生改变，每次迭代都会根据当前标签和实例相似性更新未标记实例上的标签。
- 稳定化：该过程持续进行，直到标签稳定或满足某个停止标准。
LabelPropagation 的优点：
- 利用未标记数据的信息：LabelPropagation 允许使用未标记实例中的信息来提高分类质量。当标记数据稀缺时这尤其有用。
- 对噪声的鲁棒性：该方法考虑了实例相似性并且不仅仅依赖于标签，因此可以有效地处理含有噪声的数据。
LabelPropagation 的局限性：
- 依赖于图的选择：LabelPropagation 分类的质量在很大程度上取决于图的选择和确定实例相似性的方法。错误的参数选择可能会导致不良结果。
- 计算复杂度：根据数据的大小和复杂性以及方法参数，LabelPropagation 可能需要大量的计算资源。
- 过度拟合的可能性：如果图包含太多噪声边或不正确的标签，该方法可能会过度拟合。
- 不能保证收敛：在极少数情况下，LabelPropagation 可能不会收敛到稳定的标签，需要限制迭代次数或调整其他设置。
LabelPropagation 是一种强大的方法，但它需要仔细调整参数并分析数据的图结构才能获得良好的结果。
2.29.3.1.创建 LabelPropagationClassifier 模型的代码
# Iris_LabelPropagationClassifier.py
# The code demonstrates the process of training LabelPropagation Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model.
# It also evaluates the accuracy of both the original model and the ONNX model.
# Copyright 2023, MetaQuotes Ltd.
# https://www.mql5.com
# import necessary libraries
from sklearn import datasets
from sklearn.semi_supervised import LabelPropagation
from sklearn.metrics import accuracy_score, classification_report
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as ort
import numpy as np
from sys import argv
# define the path for saving the model
data_path = argv[0]
last_index = data_path.rfind("\\") + 1
data_path = data_path[0:last_index]
# load the Iris dataset
iris = datasets.load_iris()
X = iris.data
y = iris.target
# create a LabelPropagation model
lp_model = LabelPropagation()
# train the model on the entire dataset
lp_model.fit(X, y)
# predict classes for the entire dataset
y_pred = lp_model.predict(X)
# evaluate the model's accuracy
accuracy = accuracy_score(y, y_pred)
print("Accuracy of LabelPropagation model:", accuracy)
# display the classification report
print("\nClassification Report:\n", classification_report(y, y_pred))
# define the input data type
initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))]
# export the model to ONNX format with float data type
onnx_model = convert_sklearn(lp_model, initial_types=initial_type, target_opset=12)
# save the model to a file
onnx_filename = data_path + "lp_iris.onnx"
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
# print the path to the model
print(f"Model saved to {onnx_filename}")
# load the ONNX model and make predictions
onnx_session = ort.InferenceSession(onnx_filename)
input_name = onnx_session.get_inputs()[0].name
output_name = onnx_session.get_outputs()[0].name
# display information about input tensors in ONNX
print("\nInformation about input tensors in ONNX:")
for i, input_tensor in enumerate(onnx_session.get_inputs()):
print(f"{i + 1}.Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}")
# display information about output tensors in ONNX
print("\nInformation about output tensors in ONNX:")
for i, output_tensor in enumerate(onnx_session.get_outputs()):
print(f"{i + 1}.Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}")
# convert data to floating-point format (float32)
X_float32 = X.astype(np.float32)
# predict classes for the entire dataset using ONNX
y_pred_onnx = onnx_session.run([output_name], {input_name:X_float32})[0]
# evaluate the accuracy of the ONNX model
accuracy_onnx = accuracy_score(y, y_pred_onnx)
print("\nAccuracy of LabelPropagation model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of LabelPropagation model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
模型已构建，但在转换为 ONNX 格式时出现错误。
在“错误”选项卡中，显示有关转换为 ONNX 格式的错误消息：
onnx_model = convert_sklearn(lp_model, initial_types=initial_type, target_opset=12) Iris_LabelPropagation.py 44 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.semi_supervised._label_propagation.LabelPropagation'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_LabelPropagation.py finished in 2064 ms 19 1
因此，LabelSpreading Classifier 模型无法转换为 ONNX。 onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.semi_supervised._label_propagation.LabelPropagation'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_LabelPropagation.py finished in 2064 ms 19 1
2.29.4.LabelSpreading Classifier
LabelSpreading 是一种用于分类任务的半监督学习方法。它基于在基于图形的数据结构中将标签（类）从标记实例传播到未标记实例的想法，类似于 LabelPropagation。然而，LabelSpreading 包括标签传播过程的额外稳定化和规范化。
标签传播过程：
- 它首先构建一个图，其中节点代表数据实例，节点之间的边反映实例之间的相似性或接近性。
- 初始标签分配：标记的实例被赋予了其标签，而未标记的实例则以一些未定义的标签开始。
- 图上的标签传播：根据实例之间的相似性，标记实例的标签会传播到未标记实例。
- 正则化和稳定化：LabelSpreading 包含正则化，有助于稳定标签传播过程并减少过度拟合。这是通过不仅考虑实例之间的相似性而且还考虑相邻实例的标签之间的差异来实现的。
- 迭代过程：标签可以在几次迭代中发生改变，每次迭代都会根据当前标签和正则化更新未标记实例上的标签。
- 稳定化：该过程持续进行，直到标签稳定或满足某个停止标准。
LabelSpreading 的优点：
- 利用未标记数据的信息：LabelSpreading 允许使用未标记实例中的信息来提高分类质量。
- 正则化：LabelSpreading 中正则化的存在有助于减少过度拟合，使得标签传播过程更加稳定。
LabelSpreading 的局限性：
- 依赖于图的选择：与 LabelPropagation 类似，LabelSpreading 分类的质量在很大程度上取决于图形和方法参数的选择。
- 计算复杂度：根据数据的大小和复杂性以及方法参数，LabelSpreading 可能需要大量的计算资源。
- 不总是收敛的：在极少数情况下，LabelSpreading 可能无法收敛到稳定的标签，需要限制迭代次数或调整其他设置。
LabelSpreading 是一种也需要仔细调整的方法，并且可以成为在分类任务中使用未标记数据的有力工具。
2.29.4.1.创建 LabelSpreadingClassifier 模型的代码
#Iris_LabelSpreadingClassifier.py
# The code demonstrates the process of training LabelSpreading Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model.
# It also evaluates the accuracy of both the original model and the ONNX model.
# Copyright 2023, MetaQuotes Ltd.
# import necessary libraries
from sklearn import datasets
from sklearn.semi_supervised import LabelSpreading
from sklearn.metrics import accuracy_score, classification_report
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as ort
import numpy as np
import sys
# get the script path
script_path = sys.argv[0]
last_index = script_path.rfind("\\") + 1
data_path = script_path[0:last_index]
# load the Iris dataset
iris = datasets.load_iris()
X = iris.data
y = iris.target
# create a LabelSpreading model
ls_model = LabelSpreading()
# train the model on the entire dataset
ls_model.fit(X, y)
# predict classes for the entire dataset
y_pred = ls_model.predict(X)
# evaluate the model's accuracy
accuracy = accuracy_score(y, y_pred)
print("Accuracy of LabelSpreading model:", accuracy)
# display the classification report
print("\nClassification Report:\n", classification_report(y, y_pred))
# define the input data type
initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))]
# export the model to ONNX format with float data type
onnx_model = convert_sklearn(ls_model, initial_types=initial_type,target_opset=12)
# save the model to a file
onnx_filename = data_path +“ls_iris.onnx”
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
# print the path to the model
print(f"Model saved to {onnx_filename}")
# load the ONNX model and make predictions
onnx_session = ort.InferenceSession(onnx_filename)
input_name = onnx_session.get_inputs()[0].name
output_name = onnx_session.get_outputs()[0].name
# display information about input tensors in ONNX
print("\nInformation about input tensors in ONNX:")
for i, input_tensor in enumerate(onnx_session.get_inputs()):
print(f"{i + 1}.Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}")
# display information about output tensors in ONNX
print("\nInformation about output tensors in ONNX:")
for i, output_tensor in enumerate(onnx_session.get_outputs()):
print(f"{i + 1}.Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}")
# convert data to floating-point format (float32)
X_float32 = X.astype(np.float32)
# predict classes for the entire dataset using ONNX
y_pred_onnx = onnx_session.run([output_name], {input_name:X_float32})[0]
# evaluate the accuracy of the ONNX model
accuracy_onnx = accuracy_score(y, y_pred_onnx)
print("\nAccuracy of LabelSpreading model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of LabelSpreading model:1.0
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 1.00 1.00 1.00 50
Python 2 1.00 1.00 1.00 50
Python
Python accuracy 1.00 150
Python macro avg 1.00 1.00 1.00 150
Python weighted avg 1.00 1.00 1.00 150
Python
在“错误”选项卡中，显示有关转换为 ONNX 格式的错误消息：
onnx_model = convert_sklearn(ls_model, initial_types=initial_type, target_opset=12) Iris_LabelSpreading.py 45 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.semi_supervised._label_propagation.LabelSpreading'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_LabelSpreading.py finished in 2032 ms 19 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.semi_supervised._label_propagation.LabelSpreading'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_LabelSpreading.py finished in 2032 ms 19 1
LabelSpreading Classifier 模型无法转换为 ONNX。
2.29.5.NearestCentroid Classifier
NearestCentroid （最近质心）是一种分类方法，其思想是确定每个类别的质心，并根据最近的质心对对象进行分类。该方法适用于多类问题，并且对具有线性可分类别的数据集效果良好。
NearestCentroid 过程：
- 对于每个类，都会计算一个质心，它表示属于该类的所有对象的特征的平均值。这可以通过计算该类中对象每个特征的平均值来实现。
- 当对一个新对象进行分类时，将在所有类别的质心之中计算其最近的质心。
- 将新对象分配给度量空间中质心与其最接近的类。
NearestCentroid 的优点：
- 简单且快速：NearestCentroid 是一种计算简单的方法，可以在大型数据集上快速运行。
- 适用于线性可分类别：该方法在类别线性可分或接近线性可分的任务上表现良好。
- 对于多类问题有效：NearestCentroid 适用于多类问题，并且可以用作集成中的基分类器。
NearestCentroid 的局限性：
- 对异常值的敏感性：NearestCentroid 方法对数据中的异常值很敏感，因为质心可能会因异常值的存在而严重扭曲。
- 空间偏差：如果数据中的类别具有不同的方差和形状，则 NearestCentroid 方法的效率可能会降低。
- 假设相等意味着：该方法假设类别具有大致相等的特征均值，但这在现实世界的数据中可能并不总是成立。
- 不适合非线性任务：NearestCentroid 不适用于类别之间具有非线性边界的任务。
NearestCentroid 是一种简单且可解释的分类方法，在特定场景中很有用，尤其是当类别是线性可分且数据中没有异常值时。
2.29.5.1.创建 NearestCentroid 模型的代码
# Iris_NearestCentroidClassifier.py
# The code demonstrates the process of training NearestCentroid Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model.
# It also evaluates the accuracy of both the original model and the ONNX model.
# Copyright 2023, MetaQuotes Ltd.
# https://www.mql5.com
# import necessary libraries
from sklearn import datasets
from sklearn.neighbors import NearestCentroid
from sklearn.metrics import accuracy_score, classification_report
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as ort
import numpy as np
import sys
# get the script path
script_path = sys.argv[0]
last_index = script_path.rfind("\\") + 1
data_path = script_path[0:last_index]
# load the Iris dataset
iris = datasets.load_iris()
X = iris.data
y = iris.target
# create a NearestCentroid model
nc_model = NearestCentroid()
# train the model on the entire dataset
nc_model.fit(X, y)
# predict classes for the entire dataset
y_pred = nc_model.predict(X)
# evaluate the model's accuracy
accuracy = accuracy_score(y, y_pred)
print("Accuracy of NearestCentroid model:", accuracy)
# display the classification report
print("\nClassification Report:\n", classification_report(y, y_pred))
# define the input data type
initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))]
# export the model to ONNX format with float data type
onnx_model = convert_sklearn(nc_model, initial_types=initial_type,target_opset=12)
# save the model to a file
onnx_filename = data_path +“nc_iris.onnx”
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
# print the path to the model
print(f"Model saved to {onnx_filename}")
# load the ONNX model and make predictions
onnx_session = ort.InferenceSession(onnx_filename)
input_name = onnx_session.get_inputs()[0].name
output_name = onnx_session.get_outputs()[0].name
# display information about input tensors in ONNX
print("\nInformation about input tensors in ONNX:")
for i, input_tensor in enumerate(onnx_session.get_inputs()):
print(f"{i + 1}.Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}")
# display information about output tensors in ONNX
print("\nInformation about output tensors in ONNX:")
for i, output_tensor in enumerate(onnx_session.get_outputs()):
print(f"{i + 1}.Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}")
# convert data to floating-point format (float32)
X_float32 = X.astype(np.float32)
# predict classes for the entire dataset using ONNX
y_pred_onnx = onnx_session.run([output_name], {input_name:X_float32})[0]
# evaluate the accuracy of the ONNX model
accuracy_onnx = accuracy_score(y, y_pred_onnx)
print("\nAccuracy of NearestCentroid model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of NearestCentroid model:0.9266666666666666
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.87 0.92 0.89 50
Python 2 0.91 0.86 0.89 50
Python
Python accuracy 0.93 150
Python macro avg 0.93 0.93 0.93 150
Python weighted avg 0.93 0.93 0.93 150
Python
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.87 0.92 0.89 50
Python 2 0.91 0.86 0.89 50
Python
Python accuracy 0.93 150
Python macro avg 0.93 0.93 0.93 150
Python weighted avg 0.93 0.93 0.93 150
Python
在“错误”选项卡中，显示有关转换为 ONNX 格式的错误消息：
onnx_model = convert_sklearn(nc_model, initial_types=initial_type, target_opset=12) Iris_NearestCentroid.py 45 1
onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.neighbors._nearest_centroid.NearestCentroid'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_NearestCentroid.py finished in 2131 ms 19 1
NearestCentroid Classifier 模型无法转换为 ONNX。 onnx_model = convert_topology( convert.py 208 1
topology.convert_operators(container=container, verbose=verbose) _topology.py 1532 1
self.call_shape_calculator(operator) _topology.py 1348 1
operator.infer_types() _topology.py 1163 1
raise MissingShapeCalculator( _topology.py 629 1
skl2onnx.common.exceptions.MissingShapeCalculator:Unable to find a shape calculator for type '<class 'sklearn.neighbors._nearest_centroid.NearestCentroid'>'. _topology.py 629 1
It usually means the pipeline being converted contains a _topology.py 629 1
transformer or a predictor with no corresponding converter _topology.py 629 1
implemented in sklearn-onnx.If the converted is implemented _topology.py 629 1
in another library, you need to register _topology.py 629 1
the converted so that it can be used by sklearn-onnx (function _topology.py 629 1
update_registered_converter).If the model is not yet covered _topology.py 629 1
by sklearn-onnx, you may raise an issue to _topology.py 629 1
https://github.com/onnx/sklearn-onnx/issues _topology.py 629 1
to get the converter implemented or even contribute to the _topology.py 629 1
project.If the model is a custom model, a new converter must _topology.py 629 1
be implemented.Examples can be found in the gallery. _topology.py 629 1
Iris_NearestCentroid.py finished in 2131 ms 19 1
2.29.6.Quadratic Discriminant Analysis Classifier
Quadratic Discriminant Analysis（QDA，二次判别分析） 是一种使用概率模型将数据分类的分类方法。它是线性判别分析 (LDA) 的概括，允许考虑每个类内的特征协方差。QDA 的主要思想是为每个类别建模特征分布，然后使用该分布对新对象进行分类。
QDA 流程：
- 计算每个类的分布参数，例如特征的平均值和协方差矩阵。这些参数是根据每个类别的训练数据估计的。
- 利用获得的参数，可以使用多元正态分布（或二次分布函数）计算每个类的概率密度。
- 当对一个新对象进行分类时，将计算每个类的概率密度值，并将该对象分配给概率最高的类。
二次判别分析（QDA）的优点：
- 考虑特征协方差：QDA 比 LDA 更灵活，因为它允许不同类别使用不同的协方差矩阵，从而使其更适应不同的数据结构。
- 适用于非线性边界：QDA 能够对类别之间的复杂和非线性边界进行建模。
- 对不平衡数据具有鲁棒性：QDA 在类别不平衡的任务上能表现良好。
二次判别分析（QDA）的局限性：
- 计算复杂度：QDA 需要估计每个类的参数，包括协方差矩阵，这在大型数据集上计算成本很高。
- 有限数据：当数据有限时，QDA 的效果可能会降低，并且参数估计也会变得不那么精确。
- 正态分布的假设：QDA 假设数据遵循正态分布，但对于某些类型的数据来说这可能并不成立。
- 过度拟合的风险：由于训练数据不足或特征协方差强，QDA 可能会面临过度拟合问题。
二次判别分析 (QDA) 是一种强大的分类方法，适用于各种数据类型，并且能够考虑类内的特征协方差。但是，它也存在使用时应考虑的局限性。
2.29.6.1.创建 Quadratic Discriminant Analysis 模型的代码
# Iris_QuadraticDiscriminantAnalysisClassifier.py
# The code demonstrates the process of training Quadratic Discriminant Analysis Classifier model on the Iris dataset, exporting it to ONNX format, and making predictions using the ONNX model.
# It also evaluates the accuracy of both the original model and the ONNX model.
# Copyright 2023, MetaQuotes Ltd.
# https://www.mql5.com
# import necessary libraries
from sklearn import datasets
from sklearn.discriminant_analysis import QuadraticDiscriminantAnalysis
from sklearn.metrics import accuracy_score, classification_report
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
import onnxruntime as ort
import numpy as np
import sys
# get the script path
script_path = sys.argv[0]
last_index = script_path.rfind("\\") + 1
data_path = script_path[0:last_index]
# load the Iris dataset
iris = datasets.load_iris()
X = iris.data
y = iris.target
# create a QuadraticDiscriminantAnalysis model
qda_model = QuadraticDiscriminantAnalysis()
# train the model on the entire dataset
qda_model.fit(X, y)
# predict classes for the entire dataset
y_pred = qda_model.predict(X)
# evaluate the model's accuracy
accuracy = accuracy_score(y, y_pred)
print("Accuracy of Quadratic Discriminant Analysis model:", accuracy)
# display the classification report
print("\nClassification Report:\n", classification_report(y, y_pred))
# define the input data type
initial_type = [('float_input', FloatTensorType([None, X.shape[1]]))]
# export the model to ONNX format with float data type
onnx_model = convert_sklearn(qda_model, initial_types=initial_type, target_opset=12)
# save the model to a file
onnx_filename = data_path + "qda_iris.onnx"
with open(onnx_filename, "wb") as f:
f.write(onnx_model.SerializeToString())
# print the path to the model
print(f"Model saved to {onnx_filename}")
# load the ONNX model and make predictions
onnx_session = ort.InferenceSession(onnx_filename)
input_name = onnx_session.get_inputs()[0].name
output_name = onnx_session.get_outputs()[0].name
# display information about input tensors in ONNX
print("\nInformation about input tensors in ONNX:")
for i, input_tensor in enumerate(onnx_session.get_inputs()):
print(f"{i + 1}.Name: {input_tensor.name}, Data Type: {input_tensor.type}, Shape: {input_tensor.shape}")
# display information about output tensors in ONNX
print("\nInformation about output tensors in ONNX:")
for i, output_tensor in enumerate(onnx_session.get_outputs()):
print(f"{i + 1}.Name: {output_tensor.name}, Data Type: {output_tensor.type}, Shape: {output_tensor.shape}")
# convert data to floating-point format (float32)
X_float32 = X.astype(np.float32)
# predict classes for the entire dataset using ONNX
y_pred_onnx = onnx_session.run([output_name], {input_name:X_float32})[0]
# evaluate the accuracy of the ONNX model
accuracy_onnx = accuracy_score(y, y_pred_onnx)
print("\nAccuracy of Quadratic Discriminant Analysis model in ONNX format:", accuracy_onnx)
输出：
Python Accuracy of Quadratic Discriminant Analysis model:0.98
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.96 0.97 50
Python 2 0.96 0.98 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\qda_iris.onnx
Python
Python Classification Report:
Python precision recall f1-score support
Python
Python 0 1.00 1.00 1.00 50
Python 1 0.98 0.96 0.97 50
Python 2 0.96 0.98 0.97 50
Python
Python accuracy 0.98 150
Python macro avg 0.98 0.98 0.98 150
Python weighted avg 0.98 0.98 0.98 150
Python
Python Model saved to C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\qda_iris.onnx
这次，模型成功以ONNX格式保存。但是，执行时，“错误”选项卡中会显示错误：
onnx_session = ort.InferenceSession(onnx_filename) Iris_QuadraticDiscriminantAnalysisClassifier.py 55 1
self._create_inference_session(providers, provider_options, disabled_optimizers) onnxruntime_inference_collection.py 383 1
sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model) onnxruntime_inference_collection.py 424 1
onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] :10 :INVALID_GRAPH :Load model from C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\qda_iris.onnx failed:This is an invalid mode onnxruntime_inference_collection.py 424 1
Iris_QuadraticDiscriminantAnalysisClassifier.py finished in 2063 ms 5 1
self._create_inference_session(providers, provider_options, disabled_optimizers) onnxruntime_inference_collection.py 383 1
sess = C.InferenceSession(session_options, self._model_path, True, self._read_config_from_model) onnxruntime_inference_collection.py 424 1
onnxruntime.capi.onnxruntime_pybind11_state.InvalidGraph: [ONNXRuntimeError] :10 :INVALID_GRAPH :Load model from C:\Users\user\AppData\Roaming\MetaQuotes\Terminal\D0E8209F77C8CF37AD8BF550E51FF075\MQL5\Scripts\qda_iris.onnx failed:This is an invalid mode onnxruntime_inference_collection.py 424 1
Iris_QuadraticDiscriminantAnalysisClassifier.py finished in 2063 ms 5 1
将 Quadratic Discriminant Analysis Classifier 模型转换为 ONNX 时遇到错误。
结论
在本研究中，我们利用 Scikit-learn 库版本 1.2.2，使用 Iris 数据集对 33 个分类模型进行了研究。
1.其中有 6 个模型在转换为 ONNX 格式时遇到了困难：
- DummyClassifier:Dummy Classifier（虚拟分类器）；
- GaussianProcessClassifier: Gaussian Process Classifier（高斯过程分类器）；
- LabelPropagation :Label Propagation Classifier（标签传播分类器）；
- LabelSpreading :Label Spreading Classifier（标签扩散分类器）；
- NearestCentroid:Nearest Centroid Classifier（最近质心分类器）；
- QuadraticDiscriminantAnalysis:Quadratic Discriminant Analysis Classifier（二次判别分析分类器）。
这些模型在结构和/或逻辑方面似乎更为复杂，并且它们对 ONNX 格式的适应可能需要额外的努力。它们也可能使用了不完全支持或不适合 ONNX 格式的特定数据结构或算法。
2.其余 27 个模型成功转换为 ONNX 格式，并证明其准确性得以保留。这凸显了 ONNX 作为保存和恢复机器学习模型工具的有效性，能够在不同环境和应用程序之间轻松转移模型，同时保持其性能。
成功转换为 ONNX 格式的模型完整列表包括：
- SVC：Support Vector Classifier（支持向量分类器）；
- LinearSVC：Linear Support Vector Classifier（线性支持向量分类器）；
- NuSVC：Nu Support Vector Classifier（核支持向量分类器）；
- AdaBoostClassifier：Adaptive Boosting Classifier（自适应增强分类器）；
- BaggingClassifier：Bootstrap Aggregating Classifier（自举汇聚分类器）；
- BernoulliNB：Bernoulli Naive Bayes Classifier;
- CategoricalNB:Categorical Naive Bayes Classifier（分类朴素贝叶斯分类器）；
- ComplementNB:Complement Naive Bayes Classifier;
- DecisionTreeClassifier:Decision Tree Classifier（决策树分类器）;
- ExtraTreeClassifier:Extra Tree Classifier;
- ExtraTreesClassifier:Extra Trees Classifier;
- GaussianNB:Gaussian Naive Bayes Classifier;
- GradientBoostingClassifier:Gradient Boosting Classifier（梯度提升分类器）;
- HistGradientBoostingClassifier：Histogram-Based Gradient Boosting Classifie（基于直方图的梯度提升分类器）；
- KNeighborsClassifier：k-Nearest Neighbors Classifier（k-最近邻分类器）；
- LinearDiscriminantAnalysis:Linear Discriminant Analysis Classifier;
- LogisticRegression:Logistic Regression Classifier;
- LogisticRegressionCV：Logistic Regression Classifier with Cross-Validation（带有交叉验证的逻辑回归分类器）;
- MLPClassifier：Multi-Layer Perceptron Classifier（多层感知器分类器）；
- MultinomialNB：Multinomial Naive Bayes Classifier;
- PassiveAggressiveClassifier：Passive-Aggressive Classifier;
- Perceptron：Perceptron Classifier;
- RadiusNeighborsClassifier：Radius Neighbors Classifier;
- RandomForestClassifier：Random Forest Classifier;
- RidgeClassifier：Ridge Classifier;
- RidgeClassifierCV：Ridge Classifier with Cross-Validation（带交叉验证的岭分类器）；
- SGDClassifier：Stochastic Gradient Descent Classifier（随机梯度下降分类器）。
3.此外，在研究过程中，我们确定了在 Iris 数据集上表现出优异分类性能的模型。Random Forest Classifier, Gradient Boosting Classifier, Bagging Classifier, Decision Tree Classifier, Extra Tree Classifier, Extra Trees Classifier, 和 Hist Gradient Boosting Classifier 等分类模型在预测中实现了完美的准确率。这意味着他们可以准确地确定每个鸢尾花样本所属的类别。
在为特定分类任务选择最佳模型时，这些结果尤其有价值。在鸢尾花数据上实现完美准确度的模型对于涉及类似数据的分析或分类任务而言是极佳的选择。
因此，这项研究强调了为特定任务选择正确模型的重要性，并强调了使用 ONNX 保存和应用机器学习模型进行分类任务的优势。
结论
在本文中，我们使用 Scikit-learn 版本 1.2.2 的 Iris 数据集分析了 33 个分类模型。
在我们检查的所有模型中，有六个模型很难转换为 ONNX 格式。这些模型包括 Dummy Classifier, Gaussian Process Classifier, Label Propagation Classifier, Label Spreading Classifier, Nearest Centroid Classifier, 和 Quadratic Discriminant Analysis Classifier.它们的复杂结构或逻辑可能需要额外的调整才能成功转换为 ONNX 格式。
其余 27 个模型成功转换为 ONNX 格式，并证明其准确性得以保留。这再次证明了 ONNX 在保存和恢复机器学习模型方面的效率，在保持模型性能的同时确保可移植性。
值得注意的是，一些模型，例如 Random Forest Classifier, Gradient Boosting Classifier, Bagging Classifier, Decision Tree Classifier, Extra Tree Classifier, Extra Trees Classifier 以及 Hist Gradient Boosting Classifier，在对 Iris 数据进行分类时实现了完美的准确率。这些模型对于高精度至关重要的任务尤其有吸引力。
这项研究强调了为特定任务选择正确模型的重要性，并展示了使用 ONNX 保存和在分类任务中应用机器学习模型的好处。
本文中的所有脚本也可在公共项目 “MQL5\Shared Projects\Scikit.Classification.ONNX”中找到。
本文由MetaQuotes Ltd译自俄文
原文地址： https://www.mql5.com/ru/articles/13451
附加的文件 |
Scikit-learn-ClassificationModels-ONNX.zip
(220.05 KB)
注意: MetaQuotes Ltd.将保留所有关于这些材料的权利。全部或部分复制或者转载这些材料将被禁止。
前往讨论
神经网络变得简单（第 62 部分）：在层次化模型中运用决策转换器
在最近的文章中，我们已看到了运用决策转换器方法的若干选项。该方法不仅可以分析当前状态，还可以分析先前状态的轨迹，以及在其中执行的动作。在本文中，我们将专注于在层次化模型中运用该方法。
神经网络变得简单（第 61 部分）：离线强化学习中的乐观情绪问题
在离线学习期间，我们基于训练样本数据优化了智能体的政策。成品政策令智能体对其动作充满信心。然而，这种乐观情绪并不总是正当的，并且可能会在模型操作期间导致风险增加。今天，我们要寻找降低这些风险的方法之一。
如何利用 MQL5 创建简单的多币种智能交易系统（第 3 部分）：添加交易品种、前缀和/或后缀、以及交易时段
若干交易员同事发送电子邮件或评论了如何基于经纪商提供的名称里带有前缀和/或后缀的品种使用此多币种 EA，以及如何在该多币种 EA 上实现交易时区或交易时段。
