条件GAN

!pip install -q git+https://github.com/tensorflow/docs

import keras

from keras import layers
from keras import ops
from tensorflow_docs.vis import embed
import tensorflow as tf
import numpy as np
import imageio

batch_size = 64
num_channels = 1
num_classes = 10
image_size = 28
latent_dim = 128

# 我们将使用训练集和测试集中所有可用的示例。
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
all_digits = np.concatenate([x_train, x_test])
all_labels = np.concatenate([y_train, y_test])

# 将像素值缩放到[0, 1]范围，为图像添加一个通道维度，并对标签进行独热编码。
all_digits = all_digits.astype("float32") / 255.0
all_digits = np.reshape(all_digits, (-1, 28, 28, 1))
all_labels = keras.utils.to_categorical(all_labels, 10)

# 创建tf.data.Dataset。
dataset = tf.data.Dataset.from_tensor_slices((all_digits, all_labels))
dataset = dataset.shuffle(buffer_size=1024).batch(batch_size)

print(f"训练图像的形状: {all_digits.shape}")
print(f"训练标签的形状: {all_labels.shape}")

generator_in_channels = latent_dim + num_classes
discriminator_in_channels = num_channels + num_classes
print(generator_in_channels, discriminator_in_channels)

# 创建鉴别器。
discriminator = keras.Sequential(
    [
        keras.layers.InputLayer((28, 28, discriminator_in_channels)),
        layers.Conv2D(64, (3, 3), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Conv2D(128, (3, 3), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.GlobalMaxPooling2D(),
        layers.Dense(1),
    ],
    name="discriminator",
)

# 创建生成器。
generator = keras.Sequential(
    [
        keras.layers.InputLayer((generator_in_channels,)),
        # 我们希望生成 128 + num_classes 个系数，以重塑为
        # 7x7x(128 + num_classes) 的映射。
        layers.Dense(7 * 7 * generator_in_channels),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Reshape((7, 7, generator_in_channels)),
        layers.Conv2DTranspose(128, (4, 4), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Conv2DTranspose(128, (4, 4), strides=(2, 2), padding="same"),
        layers.LeakyReLU(negative_slope=0.2),
        layers.Conv2D(1, (7, 7), padding="same", activation="sigmoid"),
    ],
    name="generator",
)

class ConditionalGAN(keras.Model):
    def __init__(self, discriminator, generator, latent_dim):
        super().__init__()
        self.discriminator = discriminator
        self.generator = generator
        self.latent_dim = latent_dim
        self.seed_generator = keras.random.SeedGenerator(1337)
        self.gen_loss_tracker = keras.metrics.Mean(name="generator_loss")
        self.disc_loss_tracker = keras.metrics.Mean(name="discriminator_loss")

    @property
    def metrics(self):
        return [self.gen_loss_tracker, self.disc_loss_tracker]

    def compile(self, d_optimizer, g_optimizer, loss_fn):
        super().compile()
        self.d_optimizer = d_optimizer
        self.g_optimizer = g_optimizer
        self.loss_fn = loss_fn

    def train_step(self, data):
        # 解压数据。
        real_images, one_hot_labels = data

        # 向标签中添加虚拟维度，以便可以与
        # 图像连接。这是给鉴别器用的。
        image_one_hot_labels = one_hot_labels[:, :, None, None]
        image_one_hot_labels = ops.repeat(
            image_one_hot_labels, repeats=[image_size * image_size]
        )
        image_one_hot_labels = ops.reshape(
            image_one_hot_labels, (-1, image_size, image_size, num_classes)
        )

        # 在潜在空间中随机采样点并连接标签。
        # 这是给生成器用的。
        batch_size = ops.shape(real_images)[0]
        random_latent_vectors = keras.random.normal(
            shape=(batch_size, self.latent_dim), seed=self.seed_generator
        )
        random_vector_labels = ops.concatenate(
            [random_latent_vectors, one_hot_labels], axis=1
        )

        # 解码噪声（根据标签引导）生成假图像。
        generated_images = self.generator(random_vector_labels)

        # 将它们与真实图像结合。注意这里我们将标签
        # 与这些图像连接。
        fake_image_and_labels = ops.concatenate(
            [generated_images, image_one_hot_labels], -1
        )
        real_image_and_labels = ops.concatenate([real_images, image_one_hot_labels], -1)
        combined_images = ops.concatenate(
            [fake_image_and_labels, real_image_and_labels], axis=0
        )

        # 汇总区分真实和假图像的标签。
        labels = ops.concatenate(
            [ops.ones((batch_size, 1)), ops.zeros((batch_size, 1))], axis=0
        )

        # 训练鉴别器。
        with tf.GradientTape() as tape:
            predictions = self.discriminator(combined_images)
            d_loss = self.loss_fn(labels, predictions)
        grads = tape.gradient(d_loss, self.discriminator.trainable_weights)
        self.d_optimizer.apply_gradients(
            zip(grads, self.discriminator.trainable_weights)
        )

        # 在潜在空间中随机采样点。
        random_latent_vectors = keras.random.normal(
            shape=(batch_size, self.latent_dim), seed=self.seed_generator
        )
        random_vector_labels = ops.concatenate(
            [random_latent_vectors, one_hot_labels], axis=1
        )

        # 汇总表示“所有真实图像”的标签。
        misleading_labels = ops.zeros((batch_size, 1))

        # 训练生成器（注意我们不应该更新
        # 鉴别器的权重）!
        with tf.GradientTape() as tape:
            fake_images = self.generator(random_vector_labels)
            fake_image_and_labels = ops.concatenate(
                [fake_images, image_one_hot_labels], -1
            )
            predictions = self.discriminator(fake_image_and_labels)
            g_loss = self.loss_fn(misleading_labels, predictions)
        grads = tape.gradient(g_loss, self.generator.trainable_weights)
        self.g_optimizer.apply_gradients(zip(grads, self.generator.trainable_weights))

        # 监控损失。
        self.gen_loss_tracker.update_state(g_loss)
        self.disc_loss_tracker.update_state(d_loss)
        return {
            "g_loss": self.gen_loss_tracker.result(),
            "d_loss": self.disc_loss_tracker.result(),
        }

cond_gan = ConditionalGAN(
    discriminator=discriminator, generator=generator, latent_dim=latent_dim
)
cond_gan.compile(
    d_optimizer=keras.optimizers.Adam(learning_rate=0.0003),
    g_optimizer=keras.optimizers.Adam(learning_rate=0.0003),
    loss_fn=keras.losses.BinaryCrossentropy(from_logits=True),
)

cond_gan.fit(dataset, epochs=20)

# 首先从我们的条件GAN中提取训练好的生成器。
trained_gen = cond_gan.generator

# 选择在插值过程中生成的中间图像数量 + 2（起始图像和最后图像）。
num_interpolation = 9  # @param {type:"integer"}

# 为插值样本生成噪声。
interpolation_noise = keras.random.normal(shape=(1, latent_dim))
interpolation_noise = ops.repeat(interpolation_noise, repeats=num_interpolation)
interpolation_noise = ops.reshape(interpolation_noise, (num_interpolation, latent_dim))


def interpolate_class(first_number, second_number):
    # 将起始和结束标签转换为独热编码向量。
    first_label = keras.utils.to_categorical([first_number], num_classes)
    second_label = keras.utils.to_categorical([second_number], num_classes)
    first_label = ops.cast(first_label, "float32")
    second_label = ops.cast(second_label, "float32")

    # 计算两个标签之间的插值向量。
    percent_second_label = ops.linspace(0, 1, num_interpolation)[:, None]
    percent_second_label = ops.cast(percent_second_label, "float32")
    interpolation_labels = (
        first_label * (1 - percent_second_label) + second_label * percent_second_label
    )

    # 将噪声和标签结合，并使用生成器进行推理。
    noise_and_labels = ops.concatenate([interpolation_noise, interpolation_labels], 1)
    fake = trained_gen.predict(noise_and_labels)
    return fake


start_class = 2  # @param {type:"slider", min:0, max:9, step:1}
end_class = 6  # @param {type:"slider", min:0, max:9, step:1}

fake_images = interpolate_class(start_class, end_class)

fake_images *= 255.0
converted_images = fake_images.astype(np.uint8)
converted_images = ops.image.resize(converted_images, (96, 96)).numpy().astype(np.uint8)
imageio.mimsave("animation.gif", converted_images[:, :, :, 0], fps=1)
embed.embed_file("animation.gif")