【学习笔记】GAN模型的学习（2）------encoder部分

1 引言
前面导入数据部分，看到迭代器就开始有点看不懂了，有时间的话出一个迭代器的专题，下面就要进入正题了generator部分，但是从作者的第一行Global generator: Encode -> quantize -> reconstruct，就开始懵逼了，为啥生成器做的是encoder的工作，我在之前看李宏毅老师的视频的时候明明generator是做decoder的工作的呀，但是不管啦，继续往下看，相信看到后面会明白的啦

2 encoder部分

    @staticmethod
    def encoder(x, config, training, C, reuse=False, actv=tf.nn.relu, scope='image'):
        """
        Process image x ([512,1024]) into a feature map of size W/16 x H/16 x C
         + C:       Bottleneck depth, controls bpp
         + Output:  Projection onto C channels, C = {2,4,8,16}
        """
        init = tf.contrib.layers.xavier_initializer()
        print('<------------ Building global {} generator architecture ------------>'.format(scope))

        def conv_block(x, filters, kernel_size=[3,3], strides=2, padding='same', actv=actv, init=init):
            bn_kwargs = {
   'center':True, 'scale':True, 'training':training, 'fused':True, 'renorm':False}
            in_kwargs = {
   'center':True, 'scale': True}
            x = tf.layers.conv2d(x, filters, kernel_size, strides=strides, padding=padding, activation=None)
            # x = tf.layers.batch_normalization(x, **bn_kwargs)
            x = tf.contrib.layers.instance_norm(x, **in_kwargs)
            x = actv(x)
            return x

        with tf.variable_scope('encoder_{}'.format(scope), reuse=reuse):

            # Run convolutions
            f = [60, 120, 240, 480, 960]
            x = tf.pad(x, [[0, 0], [