ebgan.py 文件源码

def generator(noise):
    with slim.arg_scope([slim.conv2d_transpose],
                        weights_initializer=tf.truncated_normal_initializer(stddev=0.045),
                        biases_initializer=tf.constant_initializer(value=0),
                        activation_fn=None):
        with slim.arg_scope([slim.batch_norm], is_training=train_phase, decay=0.9, epsilon=1e-5,
                            param_initializers={
                                "beta": tf.constant_initializer(value=0),
                                "gamma": tf.random_normal_initializer(mean=1, stddev=0.045)
                            }):
            weight = tf.get_variable('Generator/W', [z_dim, 2 * IMAGE_SIZE * IMAGE_SIZE], initializer=tf.truncated_normal_initializer(stddev=0.045))
            bias = tf.get_variable("Generator/b", [2 * IMAGE_SIZE * IMAGE_SIZE], initializer=tf.constant_initializer(0))

            out_1 = tf.add(tf.matmul(noise, weight, name="Generator/out_1_matmul"), bias, name="Generator/out_1_add")
            out_1 = tf.reshape(out_1, [-1, IMAGE_SIZE // 16 , IMAGE_SIZE // 16, 512], name="Generator/out_1_reshape")
            out_1 = slim.batch_norm(inputs=out_1, activation_fn=tf.nn.relu, scope="Generator/bn_1")

            out_2 = slim.conv2d_transpose(out_1, num_outputs=256, kernel_size=[5, 5], stride=2, padding="SAME", scope="Generator/deconv_2")
            out_2 = slim.batch_norm(inputs=out_2, activation_fn=tf.nn.relu, scope="Generator/bn_2")

            out_3 = slim.conv2d_transpose(out_2, num_outputs=128, kernel_size=[5, 5], stride=2, padding="SAME", scope="Generator/deconv_3")
            out_3 = slim.batch_norm(inputs=out_3, activation_fn=tf.nn.relu, scope="Generator/bn_3")

            out_4 = slim.conv2d_transpose(out_3, num_outputs=64, kernel_size=[5, 5], stride=2, padding="SAME", scope="Generator/deconv_4")
            out_4 = slim.batch_norm(inputs=out_4, activation_fn=tf.nn.relu, scope="Generator/bn_4")

            out_5 = slim.conv2d_transpose(out_4, num_outputs=3, kernel_size=[5, 5], stride=2, padding="SAME", scope="Generator/deconv_5")
            out_5 = tf.nn.tanh(out_5, name="Generator/tanh_5")

    return out_5