04、模型构建之Discriminator判别器的封装_笔记

东方耀

04、模型构建之Discriminator判别器的封装_笔记

1. 
   
2. 
   
3. def conv2d(inputs, out_channel, name, training):
   
4.     def leaky_relu(x, leak=0.2):
   
5.         return tf.maximum(x, x*leak, name='leaky_relu')
   
6.     with tf.variable_scope(name_or_scope=name):
   
7.         conv2d_output = tf.layers.conv2d(inputs, out_channel, [5, 5], (2, 2), padding='SAME')
   
8.         bn = tf.layers.batch_normalization(conv2d_output, training=training)
   
9.         return leaky_relu(bn)
   
10. 
    
11. 
    
12. class Discriminator:
    
13.     def __init__(self, channels):
    
14.         self._channels = channels
    
15.         self._reuse = False
    
16. 
    
17.     def __call__(self, inputs, training):
    
18.         # inputs shape : [N 32 32 1]
    
19.         inputs = tf.convert_to_tensor(inputs)
    
20.         with tf.variable_scope(name_or_scope='discriminator', reuse=self._reuse):
    
21.             conv2d_inputs = inputs
    
22.             for i in range(len(self._channels)):
    
23.                 conv2d_inputs = conv2d(conv2d_inputs, self._channels[i], 'conv2d-%d' % i, training)
    
24.             fc_inputs = conv2d_inputs
    
25.             with tf.variable_scope(name_or_scope='fc'):
    
26.                 flatten = tf.layers.flatten(fc_inputs, name='flatten')
    
27.                 logits = tf.layers.dense(flatten, units=2, name='logits')
    
28.         self._reuse = True
    
29.         self.variables = tf.get_collection(key=tf.GraphKeys.TRAINABLE_VARIABLES, scope='discriminator')
    
30.         return logits
    
31. 
    

复制代码

豆豆888

厉害了 我的哥