第7章 算法链电子课件

第7章算法链7.1什么是管道7.2在网格搜索使用管道7.3通用管道模型7.4模型选择与调优7.5小结7.1什么是管道

在对前几章进行学习后知道,机器学习的构建包括需要对数据集进行训练，需要进行预处理，需要特征选取，需要利用算法建立模型评估。上面四个步骤是一个流水式的工作，而管道机制实现了对全部步骤的流式化封装和管理，那么不使用管道不行吗？有些时候确实不行。假设利用make_blobs数据集生成为标准差为5的数据集。fromsklearn.datasetsimportmake_blobsfromsklearn.model_selectionimporttrain_test_splitfromsklearn.preprocessingimportStandardScalerimportmatplotlib.pyplotaspltX,y=make_blobs(n_samples=500,centers=2,cluster_std=5)X_train,X_test,y_train,y_test=train_test_split(X,y,random_state=9)scaler=StandardScaler().fit(X_train)X_train_scaled=scaler.transform(X_train)X_test_scaled=scaler.transform(X_test)plt.scatter(X_train[:,0],X_train[:,1])plt.scatter(X_train_scaled[:,0],X_train_scaled[:,1],marker='^',edgecolor='k')plt.title('trainingset&scaledtrainingset')plt.show()我们对数据进行预处理是为了算法能够更好的建立模型，加入我们使用svm算法。fromsklearn.datasetsimportmake_blobsfromsklearn.model_selectionimporttrain_test_splitfromsklearn.preprocessingimportStandardScalerfromsklearn.svmimportSVCX,y=make_blobs(n_samples=500,centers=2,cluster_std=5)X_train,X_test,y_train,y_test=train_test_split(X,y,random_state=9)scaler=StandardScaler().fit(X_train)X_train_scaled=scaler.transform(X_train)X_test_scaled=scaler.transform(X_test)svm=SVC()svm.fit(X_train_scaled,y_train)print("testscore:{}".format(svm.score(X_test_scaled,y_test)))输出结果：testscore:0.864接下来还可以通过网格搜索选取最优参数fromsklearn.datasetsimportmake_blobsfromsklearn.model_selectionimporttrain_test_splitfromsklearn.preprocessingimportStandardScalerfromsklearn.svmimportSVCfromsklearn.model_selectionimportGridSearchCVX,y=make_blobs(n_samples=500,centers=2,cluster_std=5)X_train,X_test,y_train,y_test=train_test_split(X,y,random_state=9)scaler=StandardScaler().fit(X_train)X_train_scaled=scaler.transform(X_train)X_test_scaled=scaler.transform(X_test)param_grid={'C':[0.01,0.1,1,10,100],'gamma':[0.01,0.1,1,10,100]}grid=GridSearchCV(SVC(),param_grid=param_grid,cv=5)grid.fit(X_train_scaled,y_train)print("Bestcross-validationaccuracy:{}".format(grid.best_score_))print("Bestsetscore:{}".format(grid.score(X_test_scaled,y_test)))print("Bestparameters:{}",format(grid.best_params_))输出结果Bestcross-validationaccuracy:0.9386666666666666Bestsetscore:0.952Bestparameters:{}{'C':0.1,'gamma':0.01}经过网格搜索得到的精度达到了95%，但其实上述步骤中出现了错误。我们在进行数据预处理的时候，用StandardScaler拟合了训练数据集，而后用这个拟合后的的scaler去分别转换了X_trainX_test，这一步没有错。

但在进行网格搜索的时候，用训练集来拟合的GrindSearchCV。在进行交叉验证的时候，我们把train_scaled进行了拆分，这时的train_scaled都是基于训练集进行StandardScaler拟合后再对自身进行转换，相当于我们用交叉验证中生成的测试集拟合了StandardScaler后，再用这个scaler转换这个测试集自身。从图中可以看到，这样的做法是错误的。这样一来，交叉验证的得分就是不准确的。为了解决这个问题，在交叉验证的过程中，应该在进行任何预处理之前完成数据集的划分。那怎么完成在预处理之前完成数据集的划分呢，显然如果一次一次的划分太麻烦，所以这是就需要用到管道模型。fromsklearn.datasetsimportmake_blobsfromsklearn.model_selectionimporttrain_test_splitfromsklearn.preprocessingimportStandardScalerfromsklearn.svmimportSVCfromsklearn.pipelineimportPipelineX,y=make_blobs(n_samples=500,centers=2,cluster_std=5)X_train,X_test,y_train,y_test=train_test_split(X,y,random_state=9)pipe=Pipeline([("scaler",StandardScaler()),("svm",SVC())])pipe.fit(X_train,y_train)print("Testscore:{}".format(pipe.score(X_test,y_test)))输出结果Testscore:0.904

建立管道模型后，对比我们之前属于处理以及算法建模的过程代码量大大减少。7.2在网格搜索使用管道

管道配合网格搜索使用，我们定义一个需要搜索的参数网格，并利用管道和参数网格构建一个GridSearchCV，然后需要为每个参数指定它在管道中所属的步骤。fromsklearn.datasetsimportmake_blobsfromsklearn.model_selectionimporttrain_test_splitfromsklearn.svmimportSVCfromsklearn.preprocessingimportStandardScalerfromsklearn.pipelineimportPipelinefromsklearn.model_selectionimportGridSearchCVX,y=make_blobs(n_samples=500,centers=2,cluster_std=5)X_train,X_test,y_train,y_test=train_test_split(X,y,random_state=9)param_grid={'svm__C':[0.01,0.1,1,10,100],'svm__gamma':[0.01,0.1,1,10,100]}pipe=Pipeline([("scaler",StandardScaler()),("svm",SVC())])grid=GridSearchCV(pipe,param_grid=param_grid,cv=5)grid.fit(X_train,y_train)print("Bestcross-validationaccuracy:{:.2f}".format(grid.best_score_))print("Testsetscore:{:.2f}".format(grid.score(X_test,y_test)))print("Bestparameters:{}".format(grid.best_params_))输出结果Bestcross-validationaccuracy:0.94Testsetscore:0.96Bestparameters:{'svm__C':0.1,'svm__gamma':0.1}

通过在网格搜索中使用管道，我们解决了进行任何预处理之前完成数据集的划分的问题在上述创建管道时还有一个简单的方法，就是利用make_pipeline函数fromsklearn.pipelineimportmake_pipeline#标准语法pipe=Pipeline([("scaler",StandardScaler()),("svm",SVC())])#缩写语法pipe_short=make_pipeline(StandardScaler(),SVC())7.3通用管道模型

Pipeline类不但可用于预处理和网格搜索，实际上还可以将任意数量的估计器连接在一起，还可以构建一个包含特征提取、特征选择、缩放和分类的管道。fromsklearn.preprocessingimportStandardScalerfromsklearn.preprocessingimportMinMaxScalerfromsklearn.pipelineimportmake_pipelinepipe=make_pipeline(StandardScaler(),MinMaxScaler(),StandardScaler())print("Pipelinesteps:\n{}".format(pipe.steps))Pipelinesteps:[('standardscaler-1',StandardScaler(copy=True,with_mean=True,with_std=True)),('minmaxscaler',MinMaxScaler(copy=True,feature_range=(0,1))),('standardscaler-2',StandardScaler(copy=True,with_mean=True,with_std=True))]

这里我们在管道中进行了两次预处理，因为两次预处理属于同一阶段，我们可以看到管道的工作步骤。根据需要我们可以将任意的估计器都构建到一个管道里。7.4模型选择与调优

利用管道，还可以创建一个多分类器的管道。在前面的线性回归中，我们知道容易产生过拟合，所以需要正则化。

接下来我们利用管道来对两种正则化方式进行选择以及参数的调优。fromsklearn.datasetsimportload_breast_cancerfromsklearn.model_selectionimporttrain_test_splitfromsklearn.preprocessingimportStandardScalerfromsklearn.linear_modelimportRidgefromsklearn.linear_modelimportLassofromsklearn.model_selectionimportGridSearchCVfromsklearn.pipelineimportPipelinepipe=Pipeline([('scaler',StandardScaler()),('classifier',Ridge())])param_grid=[{'classifier':[Ridge()],'scaler':[StandardScaler(),None],'classifier__alpha':[0.001,0.01,0.1,1,10,100]},{'classifier':[Lasso()],'scaler':[StandardScaler(),None],'classifier__alpha':[0.001,0.01,0.1,1,10,100]}]canner=load_breast_cancer()X_train,X_test,y_train,y_test=train_test_split(canner.data,canner.target,random_state=2)grid=GridSearchCV(pipe,