機器學習之sklearn中的pipeline

Leave a Comment / Data Science / By / August 4, 2018

如圖所示，利用pipeline我們可以方便的減少代碼量同時讓機器學習的流程變得直觀，

例如我們需要做如下操作，容易看出，訓練測試集重複了代碼，

 1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 
 vect = CountVectorizer()
 tfidf = TfidfTransformer()
 clf = SGDClassifier()

 vX = vect.fit_transform(Xtrain)
 tfidfX = tfidf.fit_transform(vX)
 predicted = clf.fit_predict(tfidfX)

 # Now evaluate all steps on test set
 vX = vect.fit_transform(Xtest)
 tfidfX = tfidf.fit_transform(vX)
 predicted = clf.fit_predict(tfidfX)

利用pipeline，上面代碼可以抽象為，

 1
 2
 3
 4
 5
 6
 7
 8
 
 pipeline = Pipeline([
 (
 (
 (
 ])
 predicted = pipeline.fit(Xtrain).predict(Xtrain)
 # Now evaluate all steps on test set
 predicted = pipeline.predict(Xtest)

注意，pipeline最後一步如果有predict()方法我們才可以對pipeline使用fit_predict()，同理，最後一步如果有transform()方法我們才可以對pipeline使用fit_transform()方法。

使用pipeline做cross validation

看如下案例，即先對輸入手寫數字的數據進行PCA降維，再通過邏輯回歸預測標籤。其中我們通過pipeline對
PCA的降維維數n_components和邏輯回歸的正則項C大小做交叉驗證，主要步驟有：

  1. 依次實例化各成分對像如
  2. 以(name, object)的tuble為元素組裝pipeline如
  3. 初始化CV參數如
  4. 實例化CV對像如
1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 
 import
 import
 from
 from
 from
 logistic = linear_model.LogisticRegression()

 pca = decomposition.PCA()
 pipe = Pipeline(steps=[(

 digits = datasets.load_digits()
 X_digits = digits.data
 y_digits = digits.target

 # Prediction
 n_components = [
 Cs = np.logspace(-

 pca.fit(X_digits)
 estimator = GridSearchCV(pipe,
 dict
 estimator.fit(X_digits, y_digits)

 plt.figure(
 plt.clf()
 plt.axes([
 plt.plot(pca.explained_variance_, linewidth=
 plt.axis(
 plt.xlabel(
 plt.ylabel(
 plt.axvline(
 estimator.best_estimator_.named_steps[
 linestyle=
 label=
 plt.legend(prop=
 plt.show()

自定義transformer

我們可以如下自定義transformer（來自

1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 
 from

 class

 def
 self.

 def
 return

 def
 return

另外，我們也可以對每個feature單獨處理，例如下面的這個比較大的流水線（來自

1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 
 pipeline = Pipeline([
 (
 (
 (
 (
 ])),
 (
 (
 (
 (
 ])),
 (
 (
 (
 (
 ])),
 (
 (
 (
 (
 (
 (
 ])),
 (
 (
 (
 (
 ]))
 ])),
 (
 (
 (
 (
 (
 (
 (
 (
 (
 ])),
 (
 ])
 
 1
 2
 3
 4
 5
 6
 7
 8
 
 class

 def
 hours = DataFrame(X[
 return

 def
 return
 
 1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 
 class

 def
 self.model = model

 def
 self.model.fit(*args, **kwargs)
 return

 def
 return

FeatureUnion

sklearn.pipeline.FeatureUnion — scikit-learn 0.19.1 documentation

 1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 
 pipeline = Pipeline([
 (
 (
 (
 (
 (
 ])),
 (
 (
 ])),
 (
 ])

整個

下面的例子中，使用FeatureUnion結合PCA降維後特徵以及選擇原特徵中的幾個作為特徵組合再餵給SVM分類，最後用grid_search 做了pca的

1
 2
 3
 4
 5
 6
 7
 8
 9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 
 from
 from
 from
 from
 from
 from

 iris = load_iris()

 X, y = iris.data, iris.target

 print

 # This dataset is way too high-dimensional. Better do PCA:
 pca = PCA()

 # Maybe some original features where good, too?
 selection = SelectKBest()

 # Build estimator from PCA and Univariate selection:

 svm = SVC(kernel=

 # Do grid search over k, n_components and C:

 pipeline = Pipeline([(
 FeatureUnion([(
 selection)])), (
 svm)])

 param_grid =
 features__pca__n_components=[
 features__univ_select__k=[
 svm__C=[

 grid_search = GridSearchCV(pipeline, param_grid=param_grid, verbose=
 grid_search.fit(X, y)

 grid_search.best_estimator_
 grid_search.best_params_
 grid_search.best_score_

Leave a Comment

Your email address will not be published.