CGAN - 条件GAN
原始GAN的缺点
代码实现
import tensorflow as tffrom tensorflow import kerasfrom tensorflow.keras import layersimport matplotlib.pyplot as plt%matplotlib inlineimport numpy as npimport glob
gpu = tf.config.experimental.list_physical_devices(device_type='GPU')tf.config.experimental.set_memory_growth(gpu[0], True)
import tensorflow.keras.datasets.mnist as mnist
(train_image, train_label), (_, _) = mnist.load_data()
train_image = train_image / 127.5 - 1
train_image = np.expand_dims(train_image, -1)
train_image.shape
dataset = tf.data.Dataset.from_tensor_slices((train_image, train_label))
AUTOTUNE = tf.data.experimental.AUTOTUNE
BATCH_SIZE = 256image_count = train_image.shape[0]noise_dim = 50
dataset = dataset.shuffle(image_count).batch(BATCH_SIZE)
def generator_model():seed = layers.Input(shape=((noise_dim,))) # 输入 形状长度为50的向量label = layers.Input(shape=(()))# 形状为空# 输入维度： 因0-9一共10个字符所以长度为10 映射成长度为50 输入序列的长度为1 x = layers.Embedding(10, 50, input_length=1)(label)#嵌入层将正整数（下标）转换为具有固定大小的向量x = layers.Flatten()(x)x = layers.concatenate([seed, x])# 与输入的seed合并x = layers.Dense(3*3*128, use_bias=False)(x)# 使用dense层转换成形状3*3通道128 的向量 不使用偏值x = layers.Reshape((3, 3, 128))(x) # reshape成3*3*128 x = layers.BatchNormalization()(x)# 批标准化x = layers.ReLU()(x) # 使用relu激活x x = layers.Conv2DTranspose(64, (3, 3), strides=(2, 2), use_bias=False)(x)# 反卷积64个卷积核 卷积核大小（3*3） 跨度2x = layers.BatchNormalization()(x)# 批标准化x = layers.ReLU()(x) #使用relu激活x # 7*7# 反卷积64个卷积核 卷积核大小（3*3） 跨度2 填充方式same 不适用偏值x = layers.Conv2DTranspose(32, (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)x = layers.BatchNormalization()(x)x = layers.ReLU()(x) # 14*14x = layers.Conv2DTranspose(1, (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)x = layers.Activation('tanh')(x)model = tf.keras.Model(inputs=[seed,label], outputs=x) # 创建模型 return model
def discriminator_model():image = tf.keras.Input(shape=((28,28,1)))label = tf.keras.Input(shape=(()))x = layers.Embedding(10, 28*28, input_length=1)(label)x = layers.Reshape((28, 28, 1))(x)x = layers.concatenate([image, x])x = layers.Conv2D(32, (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)x = layers.BatchNormalization()(x)x = layers.LeakyReLU()(x)x = layers.Dropout(0.5)(x)x = layers.Conv2D(32*2, (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)x = layers.BatchNormalization()(x)x = layers.LeakyReLU()(x)x = layers.Dropout(0.5)(x)x = layers.Conv2D(32*4, (3, 3), strides=(2, 2), padding='same', use_bias=False)(x)x = layers.BatchNormalization()(x)x = layers.LeakyReLU()(x)x = layers.Dropout(0.5)(x)x = layers.Flatten()(x)x1 = layers.Dense(1)(x)model = tf.keras.Model(inputs=[image, label], outputs=x1)return model
generator = generator_model()
discriminator = discriminator_model()
binary_cross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)category_cross_entropy = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
def discriminator_loss(real_output, fake_output):real_loss = binary_cross_entropy(tf.ones_like(real_output), real_output)fake_loss = binary_cross_entropy(tf.zeros_like(fake_output), fake_output)total_loss = real_loss + fake_lossreturn total_loss
def generator_loss(fake_output):fake_loss = binary_cross_entropy(tf.ones_like(fake_output), fake_output)return fake_loss
generator_optimizer = tf.keras.optimizers.Adam(1e-5)discriminator_optimizer = tf.keras.optimizers.Adam(1e-5)
@tf.functiondef train_step(images, labels):batchsize = labels.shape[0]noise = tf.random.normal([batchsize, noise_dim])with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:generated_images = generator((noise, labels), training=True)real_output = discriminator((images, labels), training=True)fake_output = discriminator((generated_images, labels), training=True)gen_loss = generator_loss(fake_output)disc_loss = discriminator_loss(real_output, fake_output)gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))
noise_dim = 50num = 10noise_seed = tf.random.normal([num, noise_dim])cat_seed = np.random.randint(0, 10, size=(num, 1))print(cat_seed.T)
def generate_images(model, test_noise_input, test_cat_input, epoch):print('Epoch:', epoch+1)# Notice `training` is set to False.# This is so all layers run in inference mode (batchnorm).predictions = model((test_noise_input, test_cat_input), training=False)predictions = tf.squeeze(predictions)fig = plt.figure(figsize=(10, 1))for i in range(predictions.shape[0]):plt.subplot(1, 10, i+1)plt.imshow((predictions[i, :, :] + 1)/2)plt.axis('off')# plt.savefig('image_at_epoch_{:04d}.png'.format(epoch))plt.show()
def train(dataset, epochs):for epoch in range(epochs):for image_batch, label_batch in dataset:train_step(image_batch, label_batch)if epoch%10 == 0:generate_images(generator, noise_seed, cat_seed, epoch)generate_images(generator, noise_seed, cat_seed, epoch)
EPOCHS = 200
train(dataset, EPOCHS)
