02、数据集与模型train代码(基于Keras)实现

东方耀

02、数据集与模型train代码(基于Keras)实现


使用数据集：mnist

1. # -*- coding: utf-8 -*-
   
2. __author__ = u'东方耀 微信：dfy_88888'
   
3. __date__ = '2019/11/6 14:30'
   
4. __product__ = 'PyCharm'
   
5. __filename__ = 'train_model'
   
6. from keras.layers import Input, Conv2D, MaxPooling2D, UpSampling2D, ZeroPadding2D
   
7. from keras.models import Model
   
8. from keras.callbacks import TensorBoard, EarlyStopping, ModelCheckpoint
   
9. # from keras.datasets import mnist
   
10. import numpy as np
    
11. import pandas as pd
    
12. import matplotlib.pyplot as plt
    
13. from tensorflow.examples.tutorials.mnist import input_data
    
14. 
    
15. 
    
16. # (x_train, y_train), (x_test, y_test) = mnist.load_data()
    
17. mnist = input_data.read_data_sets(train_dir='mnist_data', one_hot=False)
    
18. # (55000, 784) 已经归一化到0-1之间了 dtype=float32
    
19. x_train = mnist.train.images
    
20. # (10000, 784) 已经归一化到0-1之间了 dtype=float32
    
21. x_test = mnist.test.images
    
22. 
    
23. # 归一化 （0， 1）  MinMaxScaler
    
24. # x_train = x_train.astype('float32') / 255.
    
25. # x_test = x_test.astype('float32') / 255.
    
26. 
    
27. x_train = np.reshape(x_train, (len(x_train), 28, 28, 1))
    
28. x_test = np.reshape(x_test, (len(x_test), 28, 28, 1))
    
29. 
    
30. 
    
31. noise_factor = 0.5
    
32. x_train_noisy = x_train + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=x_train.shape)
    
33. x_test_noisy = x_test + noise_factor * np.random.normal(loc=0.0, scale=1.0, size=x_test.shape)
    
34. # 加了噪音后 不要越界
    
35. x_train_noisy = np.clip(x_train_noisy, 0., 1.)
    
36. x_test_noisy = np.clip(x_test_noisy, 0., 1.)
    
37. 
    
38. 
    
39. def plot_learning_curve(history):
    
40.     # ValueError: arrays must all be same length
    
41.     # 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
    
42.     df_history = pd.DataFrame(data=history.history)
    
43.     print(df_history)
    
44.     # print(df_history.index)
    
45.     print(df_history.columns)
    
46.     # print(df_history.dtypes)
    
47.     df_history.plot(figsize=(8, 5))
    
48.     plt.grid(True)
    
49.     # x就是DataFrame的索引
    
50.     plt.ylim(0, 0.5)
    
51.     plt.show()
    
52. 
    
53. 
    
54. def train_model():
    
55.     input_img = Input(shape=(28, 28, 1))
    
56.     x = Conv2D(16, (3, 3), activation='relu', padding='same')(input_img)
    
57.     x = MaxPooling2D((2, 2), padding='same')(x)
    
58.     x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    
59.     x = MaxPooling2D((2, 2), padding='same')(x)
    
60.     x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    
61.     encoded = MaxPooling2D((2, 2), padding='same', name='encoder')(x)
    
62. 
    
63.     # at this point the representation is (4, 4, 8) i.e. 128-dimensional 特征向量
    
64. 
    
65.     x = Conv2D(8, (3, 3), activation='relu', padding='same')(encoded)
    
66.     x = UpSampling2D((2, 2))(x)
    
67.     x = Conv2D(8, (3, 3), activation='relu', padding='same')(x)
    
68.     x = UpSampling2D((2, 2))(x)
    
69.     x = Conv2D(16, (3, 3), activation='relu', padding='valid')(x)
    
70.     x = UpSampling2D((2, 2))(x)
    
71.     # 这里激活为何用sigmoid？ 0-1之间的  自身也是0-1之间的
    
72.     decoded = Conv2D(1, (3, 3), activation='sigmoid', padding='same')(x)
    
73. 
    
74.     autoencoder = Model(inputs=input_img, outputs=decoded)
    
75.     autoencoder.summary()
    
76.     # 损失函数为何用binary_crossentropy？  图片之间的相似度越小越好？
    
77.     # binary_crossentropy 与  category_crossentropy 有何区别？
    
78.     autoencoder.compile(optimizer='adam', loss='binary_crossentropy', metrics=['mse', 'mae'])
    
79. 
    
80.     cb_tensorboard = TensorBoard(log_dir='./train_log')
    
81.     cb_earlystop = EarlyStopping(monitor='val_loss', min_delta=1e-3, patience=5)
    
82.     cb_save_best_model = ModelCheckpoint(filepath='./autoencoder_dfy.h5', monitor='val_loss', save_best_only=True)
    
83.     callbacks = [cb_tensorboard, cb_earlystop, cb_save_best_model]
    
84.     # 注意：X = x_train_noisy Y = x_train  就是无监督学习了 只是自动加了噪音数据noise
    
85.     # 5.5w =  55 000    1w = 1000 * 10
    
86.     # ValueError: If steps_per_epoch is set, the `batch_size` must be None.
    
87.     history = autoencoder.fit(x_train_noisy, x_train,
    
88.                     epochs=55,
    
89.                     batch_size=256,
    
90.                     # steps_per_epoch=55,
    
91.                     shuffle=True,
    
92.                     validation_data=(x_test_noisy, x_test),
    
93.                     # validation_steps=10,
    
94.                     validation_freq=1,
    
95.                     callbacks=callbacks)
    
96.     # list all data in history
    
97.     print(history.history.keys())
    
98.     fig, ax_array = plt.subplots(1, 2)
    
99.     ax1, ax2 = ax_array
    
100. 
     
101.     ax1.set_title('model mse')
     
102.     ax1.plot(history.history['mse'])
     
103.     ax1.plot(history.history['val_mse'])
     
104.     ax1.set_xlabel('epoch')
     
105.     ax1.set_ylabel('mse')
     
106.     ax1.legend(['train', 'validation'], loc='upper left')
     
107. 
     
108.     ax2.set_title('model loss')
     
109.     ax2.plot(history.history['loss'])
     
110.     ax2.plot(history.history['val_loss'])
     
111.     ax2.set_xlabel('epoch')
     
112.     ax2.set_ylabel('loss')
     
113.     ax2.legend(['train', 'validation'], loc='upper left')
     
114. 
     
115.     plt.show()
     
116.     plot_learning_curve(history)
     
117. 
     
118. 
     
119. train_model()
     

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】

xsoft

谢谢老师提供的资料。

sunhaichao

提供这些很不容易