03、tf-keras与keras的相同和不同及实战案例

东方耀

1. # -*- coding: utf-8 -*-
   
2. __author__ = u'东方耀 微信：dfy_88888'
   
3. __date__ = '2019/10/28 下午3:37'
   
4. __product__ = 'PyCharm'
   
5. __filename__ = 'tf_keras_classify_model'
   
6. 
   
7. import os
   
8. import sys
   
9. import time
   
10. import tensorflow as tf
    
11. import matplotlib as mpl
    
12. import matplotlib.pyplot as plt
    
13. import numpy as np
    
14. import pandas as pd
    
15. import sklearn
    
16. 
    
17. from tensorflow import keras
    
18. # tf的keras 和 单独的keras
    
19. # pip3 install keras
    
20. import keras
    
21. 
    
22. print(tf.__version__)
    
23. print(sys.version_info)
    
24. for module in mpl, np, pd, sklearn, tf, keras:
    
25.     print(module.__name__, module.__version__)
    
26. 
    
27. # https://tensorflow.google.cn/api_docs/python/tf/keras/datasets/fashion_mnist/load_data
    
28. fashion_mnist = keras.datasets.fashion_mnist
    
29. (X_train, y_train), (X_test, y_test) = fashion_mnist.load_data()
    
30. print(X_train.shape, y_train.shape)
    
31. print('测试数据集：', X_test.shape, y_test.shape)
    
32. X_valid, X_train = X_train[:5000], X_train[5000:]
    
33. y_valid, y_train = y_train[:5000], y_train[5000:]
    
34. print('验证数据集：', X_valid.shape, y_valid.shape)
    
35. print('训练数据集：', X_train.shape, y_train.shape)
    
36. 
    
37. 
    
38. def show_single_image(img_arr):
    
39.     plt.imshow(img_arr, cmap='binary')
    
40.     plt.show()
    
41. 
    
42. # show_single_image(X_train[11])
    
43. 
    
44. 
    
45. def show_multi_images(n_rows, n_cols, X_data, y_data, class_names):
    
46.     assert len(X_data) == len(y_data), '样本的特征与标签长度一致'
    
47.     assert n_rows*n_cols <= len(X_data)
    
48.     # 使用plt的子图 width, height in inches
    
49.     plt.figure(figsize=(n_cols*1.4, n_rows*1.6))
    
50.     for i in range(n_rows):
    
51.         for j in range(n_cols):
    
52.             index = i*n_cols + j
    
53.             plt.subplot(n_rows, n_cols, index+1)
    
54.             plt.imshow(X_data[index], cmap='binary', interpolation='nearest')
    
55.             plt.axis('off')
    
56.             plt.title(label=class_names[y_data[index]])
    
57.     plt.show()
    
58. 
    
59. 
    
60. class_names = ['t-shirt', 'trouser', 'pullover', 'dress',
    
61.                'coat', 'sandal', 'shirt', 'sneaker',
    
62.                'bag', 'ankle boot']
    
63. # show_multi_images(3, 5, X_train[:20], y_train[:20], class_names)
    
64. 
    
65. # tf.keras.Sequential 构建模型结构
    
66. # model = keras.Sequential(name='tf2.0_model')
    
67. # model.add(keras.layers.Flatten(input_shape=(28, 28)))
    
68. # model.add(keras.layers.Dense(units=300, activation='relu'))
    
69. # model.add(keras.layers.Dense(units=100, activation='relu'))
    
70. # model.add(keras.layers.Dense(10, activation='softmax'))
    
71. 
    
72. model = keras.Sequential(layers=[
    
73.     keras.layers.Flatten(input_shape=(28, 28)),
    
74.     keras.layers.Dense(units=300, activation='relu'),
    
75.     keras.layers.Dense(units=100, activation='relu'),
    
76.     keras.layers.Dense(10, activation='softmax')
    
77. ], name='tf2.0_model_2')
    
78. 
    
79. model.compile(optimizer=keras.optimizers.Adam(learning_rate=0.001),
    
80.               loss='sparse_categorical_crossentropy',
    
81.               metrics=['accuracy', 'mse'])
    
82. model.summary()
    
83. # print(tf.test.is_gpu_available())
    
84. 
    
85. # 开始训练
    
86. # validation_freq 是根据epochs来的 每隔多少轮进行验证
    
87. history = model.fit(x=X_train, y=y_train, epochs=10, verbose=1,
    
88.                     validation_data=(X_valid, y_valid), validation_freq=1)
    
89. print(type(history))
    
90. # 模型训练过程中的历史数据指标
    
91. print(history.history)
    
92. # validation_freq=2 的情况：
    
93. # {'loss': [2.276887303126942, 0.5237461835861206, 0.45976711897850037, 0.4327715507962487],
    
94. # 'accuracy': [0.75996363, 0.8186182, 0.8376727, 0.8458727],
    
95. # 'mse': [27.68526, 27.679394, 27.680891, 27.681673],
    
96. # 'val_loss': [0.5112990405797958, 0.45343393814563754],
    
97. # 'val_accuracy': [0.823199987411499, 0.843999981880188],
    
98. # 'val_mse': [27.655563354492188, 27.657318115234375]}
    
99. 
    
100. 
     
101. def plot_learning_curve(history):
     
102.     # ValueError: arrays must all be same length
     
103.     # 表格型数据 要求每一列的len一致 这里即：history.history字典里每个key对应的value长度一致
     
104.     df_history = pd.DataFrame(data=history.history)
     
105.     print(df_history)
     
106.     print(df_history.index)
     
107.     print(df_history.columns)
     
108.     print(df_history.dtypes)
     
109.     df_history.plot(figsize=(8, 5))
     
110.     plt.grid(True)
     
111.     # x就是DataFrame的索引
     
112.     plt.ylim(0, 1.3)
     
113.     plt.show()
     
114. 
     
115. 
     
116. plot_learning_curve(history)
     
117. 
     
118. 
     
119. 
     

复制代码

东方老师AI官网：http://www.ai111.vip
有任何问题可联系东方老师微信：dfy_88888
【微信二维码图片】

东方耀

1. from sklearn.preprocessing import StandardScaler, MinMaxScaler
   
2. # 数据的归一化处理
   
3. std_scaler = StandardScaler()
   
4. # print('训练数据集0：', X_train[0])
   
5. X_train_std = std_scaler.fit_transform(X_train.reshape(-1, 28*28))
   
6. X_train_std = X_train_std.reshape(-1, 28, 28)
   
7. print(X_train_std.shape)
   
8. X_valid_std = std_scaler.transform(X_valid.reshape(-1, 28*28))
   
9. X_valid_std = X_valid_std.reshape(-1, 28, 28)
   
10. print(X_valid_std.shape)
    
11. X_test_std = std_scaler.transform(X_test.reshape(-1, 28*28))
    
12. X_test_std = X_test_std.reshape(-1, 28, 28)

复制代码

东方耀

1. # 测试模型
   
2. test_loss, test_accuracy, test_mse = model.evaluate(X_test_std, y_test)
   
3. print(test_loss, test_accuracy, test_mse)
   
4. y_test_predict = model.predict(X_test_std, verbose=1)
   
5. print(y_test_predict.shape)
   
6. print(y_test_predict[:2])

复制代码

东方耀

1. # Callbacks: utilities called at certain points during model training.
   
2. # 实战回调函数
   
3. logdir = './callbacks'
   
4. if not os.path.exists(logdir):
   
5.     os.mkdir(logdir)
   
6. output_model_file = os.path.join(logdir, 'fashion_mnist_model.h5')
   
7. callbacks = [
   
8.     keras.callbacks.TensorBoard(logdir),
   
9.     keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=1e-3, patience=5),
   
10.     keras.callbacks.ModelCheckpoint(output_model_file, save_best_only=True)
    
11. ]
    
12. history = model.fit(x=X_train_std, y=y_train, epochs=10, verbose=1,
    
13.                     validation_data=(X_valid_std, y_valid), validation_freq=1, callbacks=callbacks)
    
14. print(type(history))
    
15. # 模型训练过程中的历史数据指标
    
16. print(history.history)

复制代码

东方耀

自带归一化功能的激活函数activation='selu'  Scaled Exponential Linear Unit (SELU)

model.add(keras.layers.Dropout(rate=0.5))
model.add(keras.layers.AlphaDropout(rate=0.5)) 更强大：1、均值和方差不变 2、归一化的性质不变
model.add(keras.layers.GaussianDropout(rate=0.5))