TensorFlow ⾃编码器（AE ）实战
第⼀部分：这⼀部分是来⾃《TensorFlow》龙良曲⽼师的书籍中的代码：import tensorflow as tf import numpy as np from PIL import Image import matplotlib .pyplot as plt from tensorflow .keras import loss ,optimizers ,Sequential
1
2
3
4
5
6#加载Fashion MNIST 图⽚数据集(train_x ,train_y ),(test_x ,test_y )=tf .keras .datats .fashion_mnist .load_data ()
1
2#归⼀化train_x ,test_x =train_x .astype (np .float32)/255.0,test_x .astype (np .float32)/255.0
送什么礼物好
1
2EPOCHES =2batch_size =64learning_rate =0.0001#只需要通过图⽚数据即可构建数据集对象，不需要标签train_db =tf .data .Datat .from_tensor_slices (train_x )train_db =train_db .shuffle (10000)train_db =train_db .batch (batch_size )# train _db =train_db .repeat (5)#构建测试集对象test_db =tf .data .Datat .from_tensor_slices (test_x )test_db =test_db .batch (batch_size )
1
2
3
4
5
6
7
8
9
10
11
12class AE (tf .keras .Model ): def __init__(lf ): super (AE , lf ).__init__() # 创建Enconder ⽹络，实现在⾃编码器类的初始化函数中 lf .encoder =Sequential ([ tf .keras .layers .Den (256), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (128), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (20) ]) # 创建Deconder ⽹络 lf .decoder =Sequential ([ tf .keras .layers .Den (128), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (256), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (784) ]) def call (lf ,inputs ,training =None ): # 前向传播函数 # 编码获得隐藏向量h ,[b ,784]->[b ,20] out =lf .encoder (inputs ) # 解码获得重建图⽚,[b ,20]->[b ,784] out_put =lf .decoder (out ) return out_put
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
进口英文19
20
21
22
23
24
25
26
#创建⽹络对象model =AE ()#指定输⼊⼤⼩model .build (input_shape =(4,784))#打印⽹络信息model .summary ()
1
2
3
4
5
6
7#创建优化器，并放置学习率optimizer =optimizers .Adam (learning_rate =learning_rate )
1
2#保存图像def Save_Image (img ,filename ): new_in = Image .new ('L', (280, 280)) index = 0 for i in range (0, 280, 28): for j in range (0, 280, 28): im = img [index ] im = Image .fromarray (im , mode ='L') new_in .paste (im , (i , j )) index += 1 new_in .save (filename )
1
2
3
4
5
6
7
8林氏
9
10
复兴遂州11
12
LOSS =[]for epoch in range (EPOCHES ): for step ,x in enumerate (train_db ): x =tf .reshape (x ,[-1,784]) with tf .GradientTape () as tape : # 前向计算获得重建的图⽚ out =model (x ) # 计算重建图⽚与输⼊之间的损失函数 loss =tf .loss .binary_crosntropy (x ,out ,from_logits =True ) loss =tf .reduce_mean (loss ) # ⾃动求导，包含两个⼦⽹络的梯度 grads =tape .gradient (loss ,model .trainable_variables ) # ⾃动更新，同时更新两个⼦⽹络 optimizer .apply_gradients (zip (grads ,model .trainable_variables )) if step %100==0: LOSS .append (float (loss )) print (epoch ,step ,float (loss )) x =next (iter (test_db )) logits =model (tf .reshape (x ,[-1,784])) x_hat =tf .sigmoid (logits ) x_hat =
tf .reshape (x_hat ,[-1,28,28]) x_concat =tf .concat ([x [:50],x_hat [:50],x_hat ],axis =0) x_concat =x_concat .numpy ()*255.0 x_concat =x_concat .astype (np .uint8) Save_Image (x_concat ,r 'E:\python 教学\图像识别\⾃编码器\images_AE\\%d.png'%epoch )plt .figure (figsize =(6,6))x =[i for i in range (len (LOSS ))]plt .plot (x ,LOSS ,label ='loss',linestyle ='-',color ='blue')plt .xlabel ('X')plt .ylabel ('loss')plt .legend ()plt .show ()
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
speak的过去式和过去分词
31
32
33
34
35
36
实验效果：以上在写代码的时候需要注意⼀个问题：就是在Save_Image 中的图像点阵图⼤⼩的问题，Save_Image中定义的步长为28，所以最后输出的图像是10*10的图像阵列，每⼀次x=next(iter(test_db))迭代得到图像是（32,28，28），意味着每⼀次的迭代得到的图像数为32，通过后⾯的拼接：x_at([x[:50],x_hat[:50]],axis=0)之后的图像为（64,28，28），虽然这个地⽅写的是x[:50]读取前50张图像，但是实际上得到的前32张，所以10*10张图像肯定是超出了索引的范围，也就是在运⾏代码的时候会出现：索引越界的错误。
所以这个地⽅该怎么改呢：可以将Save_Image中的步长设置为56，增加步长，只不过得到的图像数少⼀点，25张。将这个x_at([x[:50],x_hat[:50]],axis=0)可以修改为这个x_at([x[:15],x_hat[:15]],axis=0)。就可以了，只要索引的范围不越界就⾏。
第⼆部分：对上⾯的代码进⾏改进
import tensorflow as tf import numpy as np from PIL import Image from tensorflow .keras import loss ,optimizers ,Sequential
1
2
3
4#加载Fashion MNIST 图⽚数据集(train_x ,train_y ),(test_x ,test_y )=tf .keras .datats .fashion_mnist .load_data ()
1
2#归⼀化train_x ,test_x =train_x .astype (np .float32)/255.0,test_x .astype (np .float32)/255.0
1
2train_x =tf .reshape (train_x ,[-1,784])print (np .shape (train_x ))
1
2batch_size =64learning_rate =0.0001
1
2
encoder =Sequential ([ tf .keras .layers .Den (256), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (128), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (20)])decoder =Sequential ([ tf .keras .layers .Den (128), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (256), tf .keras .layers .Activation ('relu'), tf .keras .layers .Den (784)])autoencoder =Sequential ([ encoder , decoder ])
1
2
3
4
5
6
7
8
9
10
11
蔬菜牛肉粥
12
13
14
15
16
17
18
19
20#创建优化器，并放置学习率optimizer =optimizers .Adam (learning_rate =learning_rate )#保存图像def Save_Image (img ,filename ): new_in =Image .new ('L',(280,280)) index =0 for i in range (0,280,56): for j in range (0,280,56): im =img [index ] im =Image .fromarray (im ,mode ='L') new_in .paste (im ,(i ,j )) index +=1 new_in .save (filename )
1
2
3
4
5
6
牛奶棒7
8
9
10
11
自己的家
12
13autoencoder .compile (optimizer =optimizer ,loss =loss .binary_crosntropy )1autoencoder .fit (train_x ,train_x ,batch_size =batch_size *2,epochs =2,verbo =1)1for epoch in range (2): x =next (iter (test_db )) print (np .shape (x )) logits =model .predict (tf .reshape (x ,[-1,784])) x_hat =tf .sigmoid (logits ) x_hat =tf .reshape (x_hat ,[-1,28,28]) x_concat =tf .concat ([x [:15],x_hat [:15]],axis =0) print (np .shape (x_concat )) x_concat =x_concat .numpy ()*255.0 x_concat =x_concat .astype (np .uint8) Save_Image (x_concat ,r 'E:\python 教学\图像识别\⾃编码器\images_AE\%d.png'%epoch )1
2
3
4
5
6
7
8
9
10
11
12
