train.py 文件源码

def decoder(layers, y, BATCH_SIZE):

   layers = layers[0]
   # get all the layers from the encoder for skip connections.
   enc_conv1 = layers[0]
   enc_conv2 = layers[1]
   enc_conv3 = layers[2]
   enc_conv4 = layers[3]
   enc_conv5 = layers[4]
   enc_conv6 = layers[5]
   '''
   print 'enc_conv1:',enc_conv1
   print 'enc_conv2:',enc_conv2
   print 'enc_conv3:',enc_conv3
   print 'enc_conv4:',enc_conv4
   print 'enc_conv5:',enc_conv5
   print 'enc_conv6:',enc_conv6
   '''

   # z is the latent encoding (conv6)
   z = tcl.flatten(layers[-1])
   z = tf.concat([z,y], axis=1)
   print 'z:',z

   # reshape z to put through transpose convolutions
   s = z.get_shape().as_list()[-1]
   z = tf.reshape(z, [BATCH_SIZE, 1, 1, s])
   print 'z:',z

   dec_conv1 = tcl.convolution2d_transpose(z, 512, 4, 2, activation_fn=tf.nn.relu, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_dec_conv1')
   #dec_conv1 = tf.concat([dec_conv1, enc_conv5], axis=3)
   print 'dec_conv1:',dec_conv1

   dec_conv2 = tcl.convolution2d_transpose(dec_conv1, 512, 4, 2, activation_fn=tf.nn.relu, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_dec_conv2')
   #dec_conv2 = tf.concat([dec_conv2, enc_conv4], axis=3)
   print 'dec_conv2:',dec_conv2

   dec_conv3 = tcl.convolution2d_transpose(dec_conv2, 256, 4, 2, activation_fn=tf.nn.relu, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_dec_conv3')
   #dec_conv3 = tf.concat([dec_conv3, enc_conv3], axis=3)
   print 'dec_conv3:',dec_conv3

   dec_conv4 = tcl.convolution2d_transpose(dec_conv3, 128, 4, 2, activation_fn=tf.nn.relu, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_dec_conv4')
   #dec_conv3 = tf.concat([dec_conv4, enc_conv2], axis=3)
   print 'dec_conv4:',dec_conv4

   dec_conv5 = tcl.convolution2d_transpose(dec_conv4, 64, 4, 2, activation_fn=tf.nn.relu, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_dec_conv5')
   #dec_conv3 = tf.concat([dec_conv5, enc_conv1], axis=3)
   print 'dec_conv5:',dec_conv5

   dec_conv6 = tcl.convolution2d_transpose(dec_conv5, 3, 4, 2, activation_fn=tf.nn.relu, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_dec_conv6')
   print 'dec_conv6:',dec_conv6

   print
   print 'END G'
   print
   return dec_conv6