把它们放在一起

校验者: @片刻 翻译者: @X

模型管道化

我们已经知道一些模型可以做数据转换，一些模型可以用来预测变量。我们可以建立一个组合模型同时完成以上工作:

http://sklearn.apachecn.org/cn/0.19.0/_images/sphx_glr_plot_digits_pipe_001.png

  1. import numpy as np
  2. import matplotlib.pyplot as plt
  3. import pandas as pd
  5. from sklearn import datasets
  6. from sklearn.decomposition import PCA
  7. from sklearn.linear_model import SGDClassifier
  8. from sklearn.pipeline import Pipeline
  9. from sklearn.model_selection import GridSearchCV
  12. # Define a pipeline to search for the best combination of PCA truncation
  13. # and classifier regularization.
  14. logistic = SGDClassifier(loss='log', penalty='l2', early_stopping=True,
  15.                          max_iter=10000, tol=1e-5, random_state=0)
  16. pca = PCA()
  17. pipe = Pipeline(steps=[('pca', pca), ('logistic', logistic)])
  19. digits = datasets.load_digits()
  20. X_digits = digits.data
  21. y_digits = digits.target
  23. # Parameters of pipelines can be set using ‘__’ separated parameter names:
  24. param_grid = {
  25.     'pca__n_components': [5, 20, 30, 40, 50, 64],
  26.     'logistic__alpha': np.logspace(-4, 4, 5),
  27. }
  28. search = GridSearchCV(pipe, param_grid, iid=False, cv=5)
  29. search.fit(X_digits, y_digits)
  30. print("Best parameter (CV score=%0.3f):" % search.best_score_)
  31. print(search.best_params_)
  33. # Plot the PCA spectrum
  34. pca.fit(X_digits)
  36. fig, (ax0, ax1) = plt.subplots(nrows=2, sharex=True, figsize=(6, 6))
  37. ax0.plot(pca.explained_variance_ratio_, linewidth=2)
  38. ax0.set_ylabel('PCA explained variance')
  40. ax0.axvline(search.best_estimator_.named_steps['pca'].n_components,
  41.             linestyle=':', label='n_components chosen')

用特征面进行人脸识别

该实例用到的数据集来自 LFW_(Labeled Faces in the Wild)。数据已经进行了初步预处理

http://vis-www.cs.umass.edu/lfw/lfw-funneled.tgz (233MB)

  1. """
  2. ===================================================
  3. Faces recognition example using eigenfaces and SVMs
  4. ===================================================
  6. The dataset used in this example is a preprocessed excerpt of the
  7. "Labeled Faces in the Wild", aka LFW_:
  9.   http://vis-www.cs.umass.edu/lfw/lfw-funneled.tgz (233MB)
  11. .. _LFW: http://vis-www.cs.umass.edu/lfw/
  13. Expected results for the top 5 most represented people in the dataset:
  15. ================== ============ ======= ========== =======
  16.                    precision    recall  f1-score   support
  17. ================== ============ ======= ========== =======
  18.      Ariel Sharon       0.67      0.92      0.77        13
  19.      Colin Powell       0.75      0.78      0.76        60
  20.   Donald Rumsfeld       0.78      0.67      0.72        27
  21.     George W Bush       0.86      0.86      0.86       146
  22. Gerhard Schroeder       0.76      0.76      0.76        25
  23.       Hugo Chavez       0.67      0.67      0.67        15
  24.        Tony Blair       0.81      0.69      0.75        36
  26.       avg / total       0.80      0.80      0.80       322
  27. ================== ============ ======= ========== =======
  29. """
  30. from time import time
  31. import logging
  32. import matplotlib.pyplot as plt
  34. from sklearn.model_selection import train_test_split
  35. from sklearn.model_selection import GridSearchCV
  36. from sklearn.datasets import fetch_lfw_people
  37. from sklearn.metrics import classification_report
  38. from sklearn.metrics import confusion_matrix
  39. from sklearn.decomposition import PCA
  40. from sklearn.svm import SVC
  43. print(__doc__)
  45. # Display progress logs on stdout
  46. logging.basicConfig(level=logging.INFO, format='%(asctime)s %(message)s')
  49. # #############################################################################
  50. # Download the data, if not already on disk and load it as numpy arrays
  52. lfw_people = fetch_lfw_people(min_faces_per_person=70, resize=0.4)
  54. # introspect the images arrays to find the shapes (for plotting)
  55. n_samples, h, w = lfw_people.images.shape
  57. # for machine learning we use the 2 data directly (as relative pixel
  58. # positions info is ignored by this model)
  59. X = lfw_people.data
  60. n_features = X.shape[1]
  62. # the label to predict is the id of the person
  63. y = lfw_people.target
  64. target_names = lfw_people.target_names
  65. n_classes = target_names.shape[0]
  67. print("Total dataset size:")
  68. print("n_samples: %d" % n_samples)
  69. print("n_features: %d" % n_features)
  70. print("n_classes: %d" % n_classes)
  73. # #############################################################################
  74. # Split into a training set and a test set using a stratified k fold
  76. # split into a training and testing set
  77. X_train, X_test, y_train, y_test = train_test_split(
  78.     X, y, test_size=0.25, random_state=42)
  81. # #############################################################################
  82. # Compute a PCA (eigenfaces) on the face dataset (treated as unlabeled
  83. # dataset): unsupervised feature extraction / dimensionality reduction
  84. n_components = 150
  86. print("Extracting the top %d eigenfaces from %d faces"
  87.       % (n_components, X_train.shape[0]))
  88. t0 = time()
  89. pca = PCA(n_components=n_components, svd_solver='randomized',
  90.           whiten=True).fit(X_train)
  91. print("done in %0.3fs" % (time() - t0))
  93. eigenfaces = pca.components_.reshape((n_components, h, w))
  95. print("Projecting the input data on the eigenfaces orthonormal basis")
  96. t0 = time()
  97. X_train_pca = pca.transform(X_train)
  98. X_test_pca = pca.transform(X_test)
  99. print("done in %0.3fs" % (time() - t0))
  102. # #############################################################################
  103. # Train a SVM classification model
  105. print("Fitting the classifier to the training set")
  106. t0 = time()
  107. param_grid = {'C': [1e3, 5e3, 1e4, 5e4, 1e5],
  108.               'gamma': [0.0001, 0.0005, 0.001, 0.005, 0.01, 0.1], }
  109. clf = GridSearchCV(SVC(kernel='rbf', class_weight='balanced'),
  110.                    param_grid, cv=5, iid=False)
  111. clf = clf.fit(X_train_pca, y_train)
  112. print("done in %0.3fs" % (time() - t0))
  113. print("Best estimator found by grid search:")
  114. print(clf.best_estimator_)
  117. # #############################################################################
  118. # Quantitative evaluation of the model quality on the test set
  120. print("Predicting people's names on the test set")
  121. t0 = time()
  122. y_pred = clf.predict(X_test_pca)
  123. print("done in %0.3fs" % (time() - t0))
  125. print(classification_report(y_test, y_pred, target_names=target_names))
  126. print(confusion_matrix(y_test, y_pred, labels=range(n_classes)))
  129. # #############################################################################
  130. # Qualitative evaluation of the predictions using matplotlib
  132. def plot_gallery(images, titles, h, w, n_row=3, n_col=4):
  133.     """Helper function to plot a gallery of portraits"""
  134.     plt.figure(figsize=(1.8 * n_col, 2.4 * n_row))
  135.     plt.subplots_adjust(bottom=0, left=.01, right=.99, top=.90, hspace=.35)
  136.     for i in range(n_row * n_col):
  137.         plt.subplot(n_row, n_col, i + 1)
  138.         plt.imshow(images[i].reshape((h, w)), cmap=plt.cm.gray)
  139.         plt.title(titles[i], size=12)
  140.         plt.xticks(())
  141.         plt.yticks(())
  144. # plot the result of the prediction on a portion of the test set
  146. def title(y_pred, y_test, target_names, i):
  147.     pred_name = target_names[y_pred[i]].rsplit(' ', 1)[-1]
  148.     true_name = target_names[y_test[i]].rsplit(' ', 1)[-1]
  149.     return 'predicted: %s\ntrue:      %s' % (pred_name, true_name)
  151. prediction_titles = [title(y_pred, y_test, target_names, i)
  152.                      for i in range(y_pred.shape[0])]
  154. plot_gallery(X_test, prediction_titles, h, w)
  156. # plot the gallery of the most significative eigenfaces
  158. eigenface_titles = ["eigenface %d" % i for i in range(eigenfaces.shape[0])]
  159. plot_gallery(eigenfaces, eigenface_titles, h, w)
  161. plt.show()
prediction eigenfaces
Prediction Eigenfaces

数据集中前5名最有代表性样本的预期结果:

  1.                    precision    recall  f1-score   support
  3. Gerhard_Schroeder       0.91      0.75      0.82        28
  4.   Donald_Rumsfeld       0.84      0.82      0.83        33
  5.        Tony_Blair       0.65      0.82      0.73        34
  6.      Colin_Powell       0.78      0.88      0.83        58
  7.     George_W_Bush       0.93      0.86      0.90       129
  9.       avg / total       0.86      0.84      0.85       282

开放性问题: 股票市场结构

我们可以预测 Google 在特定时间段内的股价变动吗？

Learning a graph structure