train.py 文件源码

def encoder(x,y):

   y_dim = int(y.get_shape().as_list()[-1])

   # reshape so it's batchx1x1xy_size
   y = tf.reshape(y, shape=[BATCH_SIZE, 1, 1, y_dim])
   input_ = conv_cond_concat(x, y)

   conv1 = tcl.conv2d(input_, 64, 4, 2, activation_fn=tf.identity, normalizer_fn=tcl.batch_norm, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_enc_conv1')
   conv1 = lrelu(conv1)
   conv1 = conv_cond_concat(conv1, y)

   conv2 = tcl.conv2d(conv1, 128, 4, 2, activation_fn=tf.identity, normalizer_fn=tcl.batch_norm, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_enc_conv2')
   conv2 = lrelu(conv2)
   conv2 = conv_cond_concat(conv2, y)

   conv3 = tcl.conv2d(conv2, 256, 4, 2, activation_fn=tf.identity, normalizer_fn=tcl.batch_norm, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_enc_conv3')
   conv3 = lrelu(conv3)
   conv3 = conv_cond_concat(conv3, y)

   conv4 = tcl.conv2d(conv3, 512, 4, 2, activation_fn=tf.identity, normalizer_fn=tcl.batch_norm, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_enc_conv4')
   conv4 = lrelu(conv4)
   conv4 = conv_cond_concat(conv4, y)

   conv5 = tcl.conv2d(conv4, 512, 4, 2, activation_fn=tf.identity, normalizer_fn=tcl.batch_norm, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_enc_conv5')
   conv5 = lrelu(conv5)
   conv5 = conv_cond_concat(conv5, y)

   conv6 = tcl.conv2d(conv5, 512, 4, 2, activation_fn=tf.identity, normalizer_fn=tcl.batch_norm, weights_initializer=tf.random_normal_initializer(stddev=0.02), scope='g_enc_conv6')
   conv6 = lrelu(conv6)

   print 'conv1:',conv1
   print 'conv2:',conv2
   print 'conv3:',conv3
   print 'conv4:',conv4
   print 'conv5:',conv5
   print 'conv6:',conv6
   out = [conv1, conv2, conv3, conv4, conv5, conv6]
   return out,y