Restricted Boltzmann Machine features for digit classification

http://scikit-learn.org/stable/auto_examples/neural_networks/plot_rbm_logistic_classification.html#sphx-glr-auto-examples-neural-networks-plot-rbm-logistic-classification-py

此範例將使用BernoulliRBM特徵選取方法，提升手寫數字識別的精確率，伯努利限制玻爾茲曼機器模型（`BernoulliRBM

`）將可以對數據做有效的非線性
特徵提取的處理。
為了讓此模型訓練出來更為強健，將輸入的圖檔，分別做上左右下，一像素的平移，用以增加更多訓練資料，
訓練網路的參數是使用grid search演算法，但此訓練太耗費時間，因此不再這重現，。
此範例結果將比較，
1.使用原本的像素值做的邏輯回歸
2.使用BernoulliRBM做特徵選取的邏輯回歸
結果將顯示:使用BernoulliRBM將可以提升分類的準確度。

(一)引入函式庫與資料

  1. from __future__ import print_function
  3. print(__doc__)
  5. # Authors: Yann N. Dauphin, Vlad Niculae, Gabriel Synnaeve
  6. # License: BSD
  8. import numpy as np
  9. import matplotlib.pyplot as plt
  11. from scipy.ndimage import convolve
  12. from sklearn import linear_model, datasets, metrics
  13. from sklearn.model_selection import train_test_split
  14. from sklearn.neural_network import BernoulliRBM
  15. from sklearn.pipeline import Pipeline

(二)資料前處理、讀取資料、選取模型

  1. def nudge_dataset(X, Y):
  2.     """
  3.     此副函式是用來將輸入資料的數字圖形，分別做上左右下一像素的平移，目的是製造更多的訓練資料讓模型訓練出來更強健
  4.     """
  5.     direction_vectors = [
  6.         [[0, 1, 0],
  7.          [0, 0, 0],
  8.          [0, 0, 0]],
  10.         [[0, 0, 0],
  11.          [1, 0, 0],
  12.          [0, 0, 0]],
  14.         [[0, 0, 0],
  15.          [0, 0, 1],
  16.          [0, 0, 0]],
  18.         [[0, 0, 0],
  19.          [0, 0, 0],
  20.          [0, 1, 0]]]
  22.     shift = lambda x, w: convolve(x.reshape((8, 8)), mode='constant',
  23.                                   weights=w).ravel()
  24.     X = np.concatenate([X] +
  25.                        [np.apply_along_axis(shift, 1, X, vector)
  26.                         for vector in direction_vectors])
  27.     Y = np.concatenate([Y for _ in range(5)], axis=0)
  28.     return X, Y
  30. # Load Data
  31. digits = datasets.load_digits()
  32. X = np.asarray(digits.data, 'float32')
  33. X, Y = nudge_dataset(X, digits.target)
  34. X = (X - np.min(X, 0)) / (np.max(X, 0) + 0.0001)  # 將灰階影像降尺度降到[0,1]
  35. # 將資料切割成訓練集與測試集
  36. X_train, X_test, Y_train, Y_test = train_test_split(X, Y,
  37.                                                     test_size=0.2,
  38.                                                     random_state=0)
  40. # Models we will use
  41. logistic = linear_model.LogisticRegression()
  42. rbm = BernoulliRBM(random_state=0, verbose=True)
  44. classifier = Pipeline(steps=[('rbm', rbm), ('logistic', logistic)])

(三)設定模型參數與訓練模型

  1. # 參數選擇需使用cross-validation去比較
  2. # 此參數是使用GridSearchCV找出來的. Here we are not performing cross-validation to save time.
  3. #GridSratch 就是將參數設定好，跑過全部參數後去找結果最好的一組參數
  4. rbm.learning_rate = 0.06
  5. rbm.n_iter = 20
  6. #.n_components = 100 表示隱藏層單元為100，即表示萃取出100個特徵，特徵萃取的越多準確率會越高，但越耗時間
  7. rbm.n_components = 100
  8. logistic.C = 6000.0
  10. # Training RBM-Logistic Pipeline
  11. classifier.fit(X_train, Y_train)
  13. # Training Logistic regression
  14. logistic_classifier = linear_model.LogisticRegression(C=100.0)
  15. logistic_classifier.fit(X_train, Y_train)

(四)評估模型的分辨準確率

  1. print()
  2. print("Logistic regression using RBM features:\n%s\n" % (
  3.     metrics.classification_report(
  4.         Y_test,
  5.         classifier.predict(X_test))))
  7. print("Logistic regression using raw pixel features:\n%s\n" % (
  8.     metrics.classification_report(
  9.         Y_test,
  10.         logistic_classifier.predict(X_test))))

Ex 2: Restricted Boltzmann Machine features for digit classification - 图1

圖1:使用RBM演算法後準確率為0.95

Ex 2: Restricted Boltzmann Machine features for digit classification - 图2

圖2:不使用任何特徵選取方法做的做的邏輯回歸準確率0.77

(五)畫出100個RBM萃取出的特徵

  1. plt.figure(figsize=(4.2, 4))
  2. for i, comp in enumerate(rbm.components_):
  3.     plt.subplot(10, 10, i + 1)
  4.     plt.imshow(comp.reshape((8, 8)), cmap=plt.cm.gray_r,
  5.                interpolation='nearest')
  6.     plt.xticks(())
  7.     plt.yticks(())
  8. plt.suptitle('100 components extracted by RBM', fontsize=16)
  9. plt.subplots_adjust(0.08, 0.02, 0.92, 0.85, 0.08, 0.23)
  11. plt.show()

Ex 2: Restricted Boltzmann Machine features for digit classification - 图3

圖3:使用RBM演算法，尋找出來的特徵

(六)完整程式碼

  1. from __future__ import print_function
  3. print(__doc__)
  5. # Authors: Yann N. Dauphin, Vlad Niculae, Gabriel Synnaeve
  6. # License: BSD
  8. import numpy as np
  9. import matplotlib.pyplot as plt
  11. from scipy.ndimage import convolve
  12. from sklearn import linear_model, datasets, metrics
  13. from sklearn.model_selection import train_test_split
  14. from sklearn.neural_network import BernoulliRBM
  15. from sklearn.pipeline import Pipeline
  18. ###############################################################################
  19. # Setting up
  21. def nudge_dataset(X, Y):
  22.     """
  23.     This produces a dataset 5 times bigger than the original one,
  24.     by moving the 8x8 images in X around by 1px to left, right, down, up
  25.     """
  26.     direction_vectors = [
  27.         [[0, 1, 0],
  28.          [0, 0, 0],
  29.          [0, 0, 0]],
  31.         [[0, 0, 0],
  32.          [1, 0, 0],
  33.          [0, 0, 0]],
  35.         [[0, 0, 0],
  36.          [0, 0, 1],
  37.          [0, 0, 0]],
  39.         [[0, 0, 0],
  40.          [0, 0, 0],
  41.          [0, 1, 0]]]
  43.     shift = lambda x, w: convolve(x.reshape((8, 8)), mode='constant',
  44.                                   weights=w).ravel()
  45.     X = np.concatenate([X] +
  46.                        [np.apply_along_axis(shift, 1, X, vector)
  47.                         for vector in direction_vectors])
  48.     Y = np.concatenate([Y for _ in range(5)], axis=0)
  49.     return X, Y
  51. # Load Data
  52. digits = datasets.load_digits()
  53. X = np.asarray(digits.data, 'float32')
  54. X, Y = nudge_dataset(X, digits.target)
  55. X = (X - np.min(X, 0)) / (np.max(X, 0) + 0.0001)  # 0-1 scaling
  57. X_train, X_test, Y_train, Y_test = train_test_split(X, Y,
  58.                                                     test_size=0.2,
  59.                                                     random_state=0)
  61. # Models we will use
  62. logistic = linear_model.LogisticRegression()
  63. rbm = BernoulliRBM(random_state=0, verbose=True)
  65. classifier = Pipeline(steps=[('rbm', rbm), ('logistic', logistic)])
  67. ###############################################################################
  68. # Training
  70. # Hyper-parameters. These were set by cross-validation,
  71. # using a GridSearchCV. Here we are not performing cross-validation to
  72. # save time.
  73. rbm.learning_rate = 0.06
  74. rbm.n_iter = 20
  75. # More components tend to give better prediction performance, but larger
  76. # fitting time
  77. rbm.n_components = 100
  78. logistic.C = 6000.0
  80. # Training RBM-Logistic Pipeline
  81. classifier.fit(X_train, Y_train)
  83. # Training Logistic regression
  84. logistic_classifier = linear_model.LogisticRegression(C=100.0)
  85. logistic_classifier.fit(X_train, Y_train)
  87. ###############################################################################
  88. # Evaluation
  90. print()
  91. print("Logistic regression using RBM features:\n%s\n" % (
  92.     metrics.classification_report(
  93.         Y_test,
  94.         classifier.predict(X_test))))
  96. print("Logistic regression using raw pixel features:\n%s\n" % (
  97.     metrics.classification_report(
  98.         Y_test,
  99.         logistic_classifier.predict(X_test))))
  101. ###############################################################################
  102. # Plotting
  104. plt.figure(figsize=(4.2, 4))
  105. for i, comp in enumerate(rbm.components_):
  106.     plt.subplot(10, 10, i + 1)
  107.     plt.imshow(comp.reshape((8, 8)), cmap=plt.cm.gray_r,
  108.                interpolation='nearest')
  109.     plt.xticks(())
  110.     plt.yticks(())
  111. plt.suptitle('100 components extracted by RBM', fontsize=16)
  112. plt.subplots_adjust(0.08, 0.02, 0.92, 0.85, 0.08, 0.23)
  114. plt.show()