with tf.variable_scope('A'):
    inputs = Conv2DNormLReLU(inputs, 32, 7)
    inputs = Conv2DNormLReLU(inputs, 64, strides=2)
    inputs = Conv2DNormLReLU(inputs, 64)

with tf.variable_scope('B'):
    inputs = Conv2DNormLReLU(inputs, 128, strides=2)
    inputs = Conv2DNormLReLU(inputs, 128)

with tf.variable_scope('C'):
    inputs = Conv2DNormLReLU(inputs, 128)
    inputs = self.InvertedRes_block(inputs, 2, 256, 1, 'r1')
    inputs = self.InvertedRes_block(inputs, 2, 256, 1, 'r2')
    inputs = self.InvertedRes_block(inputs, 2, 256, 1, 'r3')
    inputs = self.InvertedRes_block(inputs, 2, 256, 1, 'r4')
    inputs = Conv2DNormLReLU(inputs, 128)

with tf.variable_scope('D'):
    inputs = Unsample(inputs, 128)
    inputs = Conv2DNormLReLU(inputs, 128)

with tf.variable_scope('E'):
    inputs = Unsample(inputs,64)
    inputs = Conv2DNormLReLU(inputs, 64)
    inputs = Conv2DNormLReLU(inputs, 32, 7)
with tf.variable_scope('out_layer'):
    out = Conv2D(inputs, filters =3, kernel_size=1, strides=1)
    self.fake = tf.tanh(out)

def D_net(x_init,ch, n_dis,sn, scope, reuse):
    channel = ch // 2
    with tf.variable_scope(scope, reuse=reuse):
        x = conv(x_init, channel, kernel=3, stride=1, pad=1, use_bias=False, sn=sn, scope='conv_0')
        x = lrelu(x, 0.2)

        for i in range(1, n_dis):
            x = conv(x, channel * 2, kernel=3, stride=2, pad=1, use_bias=False, sn=sn, scope='conv_s2_' + str(i))
            x = lrelu(x, 0.2)

            x = conv(x, channel * 4, kernel=3, stride=1, pad=1, use_bias=False, sn=sn, scope='conv_s1_' + str(i))
            x = layer_norm(x, scope='1_norm_' + str(i))
            x = lrelu(x, 0.2)

            channel = channel * 2

        x = conv(x, channel * 2, kernel=3, stride=1, pad=1, use_bias=False, sn=sn, scope='last_conv')
        x = layer_norm(x, scope='2_ins_norm')
        x = lrelu(x, 0.2)

        x = conv(x, channels=1, kernel=3, stride=1, pad=1, use_bias=False, sn=sn, scope='D_logit')

        return x

def con_sty_loss(vgg, real, anime, fake):

    # 真實圖片向量化
    vgg.build(real)
    real_feature_map = vgg.conv4_4_no_activation

    # 生成圖片向量化
    vgg.build(fake)
    fake_feature_map = vgg.conv4_4_no_activation

    # 漫畫風(fēng)格向量化
    vgg.build(anime[:fake_feature_map.shape[0]])
    anime_feature_map = vgg.conv4_4_no_activation

    # 真實圖片與生成圖片的損失
    c_loss = L1_loss(real_feature_map, fake_feature_map)
    # 漫畫風(fēng)格與生成圖片的損失
    s_loss = style_loss(anime_feature_map, fake_feature_map)

    return c_loss, s_loss

c_loss, s_loss = con_sty_loss(self.vgg, self.real, self.anime_gray, self.generated)
t_loss = self.con_weight * c_loss + self.sty_weight * s_loss + color_loss(self.real,self.generated) * self.color_weight + tv_loss

python train.py --dataset Hayao --epoch 101 --init_epoch 10