began_model.py 文件源码

def decoder(self, z, embedding, reuse=None):
        with tf.variable_scope("decoder", reuse=reuse):
            with slim.arg_scope([slim.conv2d, slim.fully_connected],
                                weights_initializer=tf.contrib.layers.variance_scaling_initializer(),
                                weights_regularizer=slim.l2_regularizer(5e-4),
                                bias_initializer=tf.zeros_initializer()):
                with slim.arg_scope([slim.conv2d], padding="SAME",
                                    activation_fn=tf.nn.elu, stride=1):
                    x = slim.fully_connected(z, 8 * 8 * embedding, activation_fn=None)
                    x = tf.reshape(x, [-1, 8, 8, embedding])

                    for i in range(self.conv_repeat_num):
                        x = slim.repeat(x, 2, slim.conv2d, embedding, 3)
                        if i < self.conv_repeat_num - 1:
                            x = resize_nn(x, 2)  # NN up-sampling

                    x = slim.conv2d(x, 3, 3, activation_fn=None)
        return x