前一天我們講完了GAN的核心概念，今天要來用python去實作看看GAN囉！我們這次的程式碼是參考 https://keras.io/examples/generative/cyclegan/ 這篇文章的，由於程式碼過長，因此筆者只會拿部分重要程式碼來討論，若是想要完整程式碼，可以參考網站看看。

首先第一步要先下載這次要做訓練的資料集，我們這次用到了tensorflow裡的horse2zebra資料集。筆者在直接用程式碼下載資料的時候遇到了點問題，如果讀者也遇到相同問題無法下載，可以到 https://people.eecs.berkeley.edu/%7Etaesung_park/CycleGAN/datasets/ 直接下載資料集，下載後放到 C:\Users\使用者名稱\tensor_flow_datasets\cycle_gan\horse2zebra。

dataset, _ = tfds.load("cycle_gan/horse2zebra", with_info=True, as_supervised=True)
train_horses, train_zebras = dataset["trainA"], dataset["trainB"]
test_horses, test_zebras = dataset["testA"], dataset["testB"]

接下來要來建構GAN模型，首先我們先建立好GAN模型『生成』的部分。

def get_resnet_generator(
    filters=64,
    num_downsampling_blocks=2,
    num_residual_blocks=9,
    num_upsample_blocks=2,
    gamma_initializer=gamma_init,
    name=None,
):
    img_input = layers.Input(shape=input_img_size, name=name + "_img_input")
    x = ReflectionPadding2D(padding=(3, 3))(img_input)
    x = layers.Conv2D(filters, (7, 7), kernel_initializer=kernel_init, use_bias=False)(
        x
    )
    x = tfa.layers.InstanceNormalization(gamma_initializer=gamma_initializer)(x)
    x = layers.Activation("relu")(x)

    # Downsampling
    for _ in range(num_downsampling_blocks):
        filters *= 2
        x = downsample(x, filters=filters, activation=layers.Activation("relu"))

    # Residual blocks
    for _ in range(num_residual_blocks):
        x = residual_block(x, activation=layers.Activation("relu"))

    # Upsampling
    for _ in range(num_upsample_blocks):
        filters //= 2
        x = upsample(x, filters, activation=layers.Activation("relu"))

    # Final block
    x = ReflectionPadding2D(padding=(3, 3))(x)
    x = layers.Conv2D(3, (7, 7), padding="valid")(x)
    x = layers.Activation("tanh")(x)

    model = keras.models.Model(img_input, x, name=name)
    return model

生成的部分完成過後，我們要來做discriminator，也就是評估生成資料和真實資料的部分。

def get_discriminator(
    filters=64, kernel_initializer=kernel_init, num_downsampling=3, name=None
):
    img_input = layers.Input(shape=input_img_size, name=name + "_img_input")
    x = layers.Conv2D(
        filters,
        (4, 4),
        strides=(2, 2),
        padding="same",
        kernel_initializer=kernel_initializer,
    )(img_input)
    x = layers.LeakyReLU(0.2)(x)

    num_filters = filters
    for num_downsample_block in range(3):
        num_filters *= 2
        if num_downsample_block < 2:
            x = downsample(
                x,
                filters=num_filters,
                activation=layers.LeakyReLU(0.2),
                kernel_size=(4, 4),
                strides=(2, 2),
            )
        else:
            x = downsample(
                x,
                filters=num_filters,
                activation=layers.LeakyReLU(0.2),
                kernel_size=(4, 4),
                strides=(1, 1),
            )

    x = layers.Conv2D(
        1, (4, 4), strides=(1, 1), padding="same", kernel_initializer=kernel_initializer
    )(x)

    model = keras.models.Model(inputs=img_input, outputs=x, name=name)
    return model


# Get the generators
gen_G = get_resnet_generator(name="generator_G")
gen_F = get_resnet_generator(name="generator_F")

# Get the discriminators
disc_X = get_discriminator(name="discriminator_X")
disc_Y = get_discriminator(name="discriminator_Y")

生成和discriminator都完成過後就可以將他們給合在一起、正式建構GAN模型啦！

class CycleGan(keras.Model):
    def __init__(
        self,
        generator_G,
        generator_F,
        discriminator_X,
        discriminator_Y,
        lambda_cycle=10.0,
        lambda_identity=0.5,
    ):
        super().__init__()
        self.gen_G = generator_G
        self.gen_F = generator_F
        self.disc_X = discriminator_X
        self.disc_Y = discriminator_Y
        self.lambda_cycle = lambda_cycle
        self.lambda_identity = lambda_identity

    def compile(
        self,
        gen_G_optimizer,
        gen_F_optimizer,
        disc_X_optimizer,
        disc_Y_optimizer,
        gen_loss_fn,
        disc_loss_fn,
    ):
        super().compile()
        self.gen_G_optimizer = gen_G_optimizer
        self.gen_F_optimizer = gen_F_optimizer
        self.disc_X_optimizer = disc_X_optimizer
        self.disc_Y_optimizer = disc_Y_optimizer
        self.generator_loss_fn = gen_loss_fn
        self.discriminator_loss_fn = disc_loss_fn
        self.cycle_loss_fn = keras.losses.MeanAbsoluteError()
        self.identity_loss_fn = keras.losses.MeanAbsoluteError()

    def train_step(self, batch_data):
        # x is Horse and y is zebra
        real_x, real_y = batch_data

        with tf.GradientTape(persistent=True) as tape:
            # Horse to fake zebra
            fake_y = self.gen_G(real_x, training=True)
            # Zebra to fake horse -> y2x
            fake_x = self.gen_F(real_y, training=True)

            # Cycle (Horse to fake zebra to fake horse): x -> y -> x
            cycled_x = self.gen_F(fake_y, training=True)
            # Cycle (Zebra to fake horse to fake zebra) y -> x -> y
            cycled_y = self.gen_G(fake_x, training=True)

            # Identity mapping
            same_x = self.gen_F(real_x, training=True)
            same_y = self.gen_G(real_y, training=True)

            # Discriminator output
            disc_real_x = self.disc_X(real_x, training=True)
            disc_fake_x = self.disc_X(fake_x, training=True)

            disc_real_y = self.disc_Y(real_y, training=True)
            disc_fake_y = self.disc_Y(fake_y, training=True)

            # Generator adversarial loss
            gen_G_loss = self.generator_loss_fn(disc_fake_y)
            gen_F_loss = self.generator_loss_fn(disc_fake_x)

            # Generator cycle loss
            cycle_loss_G = self.cycle_loss_fn(real_y, cycled_y) * self.lambda_cycle
            cycle_loss_F = self.cycle_loss_fn(real_x, cycled_x) * self.lambda_cycle

            # Generator identity loss
            id_loss_G = (
                self.identity_loss_fn(real_y, same_y)
                * self.lambda_cycle
                * self.lambda_identity
            )
            id_loss_F = (
                self.identity_loss_fn(real_x, same_x)
                * self.lambda_cycle
                * self.lambda_identity
            )

            # Total generator loss
            total_loss_G = gen_G_loss + cycle_loss_G + id_loss_G
            total_loss_F = gen_F_loss + cycle_loss_F + id_loss_F

            # Discriminator loss
            disc_X_loss = self.discriminator_loss_fn(disc_real_x, disc_fake_x)
            disc_Y_loss = self.discriminator_loss_fn(disc_real_y, disc_fake_y)

        # Get the gradients for the generators
        grads_G = tape.gradient(total_loss_G, self.gen_G.trainable_variables)
        grads_F = tape.gradient(total_loss_F, self.gen_F.trainable_variables)

        # Get the gradients for the discriminators
        disc_X_grads = tape.gradient(disc_X_loss, self.disc_X.trainable_variables)
        disc_Y_grads = tape.gradient(disc_Y_loss, self.disc_Y.trainable_variables)

        # Update the weights of the generators
        self.gen_G_optimizer.apply_gradients(
            zip(grads_G, self.gen_G.trainable_variables)
        )
        self.gen_F_optimizer.apply_gradients(
            zip(grads_F, self.gen_F.trainable_variables)
        )

        # Update the weights of the discriminators
        self.disc_X_optimizer.apply_gradients(
            zip(disc_X_grads, self.disc_X.trainable_variables)
        )
        self.disc_Y_optimizer.apply_gradients(
            zip(disc_Y_grads, self.disc_Y.trainable_variables)
        )

        return {
            "G_loss": total_loss_G,
            "F_loss": total_loss_F,
            "D_X_loss": disc_X_loss,
            "D_Y_loss": disc_Y_loss,
        }

建構完成後就會進入漫長的訓練過程，筆者這邊大概訓練了2個小時左右。如果讀者們想要趕快看到結果，可以嘗試將learning rate給提高一些，雖然這樣會導致結果可能沒有那麼的好，這邊是讀者們需要自己去評估、衡量的部分。

cycle_gan_model.fit(
    tf.data.Dataset.zip((train_horses, train_zebras)),
    epochs=1,
    callbacks=[plotter, model_checkpoint_callback],
)

完成後就可以看到GAN模型所創造出來的『仿品』囉！

參考資料：https://keras.io/examples/generative/cyclegan/