学习笔记TF051:生成式对抗网络

生成式对抗网络(gennerative adversarial network,GAN)，谷歌2014年提出网络模型。灵感自二人博弈的零和博弈，目前最火的非监督深度学习。GAN之父，Ian J.Goodfellow，公认人工智能顶级专家。

原理。
生成式对搞网络包含一个生成模型(generative model,G)和一个判别模型(discriminative model,D)。Ian J.Goodfellow、Jean Pouget-Abadie、Mehdi Mirza、Bing Xu、David Warde-Farley、Sherjil Ozair、Aaron Courville、Yoshua Bengio论文，《Generative Adversarial Network》，https://arxiv.org/abs/1406.2661 。
生成式对抗网络结构：
噪声数据->生成模型->假图片---|
|->判别模型->真/假
打乱训练数据->训练集->真图片-|
生成式对抗网络主要解决如何从训练样本中学习出新样本。生成模型负责训练出样本的分布，如果训练样本是图片就生成相似的图片，如果训练样本是文章名子就生成相似的文章名子。判别模型是一个二分类器，用来判断输入样本是真实数据还是训练生成的样本。
生成式对抗网络优化，是一个二元极小极大博弈(minimax two-player game)问题。使生成模型输出在输入给判别模型时，判断模型秀难判断是真实数据还是虚似数据。训练好的生成模型，能把一个噪声向量转化成和训练集类似的样本。Argustus Odena、Christopher Olah、Jonathon Shlens论文《Coditional Image Synthesis with Auxiliary Classifier GANs》。
辅助分类器生成式对抗网络(auxiliary classifier GAN，AC-GAN)实现。

生成式对抗网络应用。生成数字，生成人脸图像。

生成式对抗网络实现。https://github.com/fchollet/keras/blob/master/examples/mnist_acgan.py 。
Augustus Odena、Chistopher Olah和Jonathon Shlens 论文《Conditional Image Synthesis With Auxiliary Classifier GANs》。
通过噪声，让生成模型G生成虚假数据，和真实数据一起送到判别模型D，判别模型一方面输出数据真/假,一方面输出图片分类。
首先定义生成模型，目的是生成一对(z,L)数据，z是噪声向量，L是(1,28,28)的图像空间。

def build_generator(latent_size):
    cnn = Sequential()
    cnn.add(Dense(1024, input_dim=latent_size, activation='relu'))
    cnn.add(Dense(128 * 7 * 7, activation='relu'))
    cnn.add(Reshape((128, 7, 7)))
    #上采样，图你尺寸变为 14X14
    cnn.add(UpSampling2D(size=(2,2)))
    cnn.add(Convolution2D(256, 5, 5, border_mode='same', activation='relu', init='glorot_normal'))
    #上采样，图像尺寸变为28X28
    cnn.add(UpSampling2D(size=(2,2)))
    cnn.add(Convolution2D(128, 5, 5, border_mode='same', activation='relu', init='glorot_normal'))
    #规约到1个通道
    cnn.add(Convolution2D(1, 2, 2, border_mode='same', activation='tanh', init='glorot_normal'))
    #生成模型输入层,特征向量
    latent = Input(shape=(latent_size, ))
    #生成模型输入层，标记
    image_class = Input(shape=(1,), dtype='int32')
    cls = Flatten()(Embedding(10, latent_size, init='glorot_normal')(image_class))
    h = merge([latent, cls], mode='mul')
    fake_image = cnn(h) #输出虚假图片
    return Model(input=[latent, image_class], output=fake_image)

定义判别模型，输入(1,28,28)图片，输出两个值，一个是判别模型认为这张图片是否是虚假图片，另一个是判别模型认为这第图片所属分类。

def build_discriminator();
    #采用激活函数Leaky ReLU来替换标准的卷积神经网络中的激活函数
    cnn = Wequential()
    cnn.add(Convolution2D(32, 3, 3, border_mode='same', subsample=(2, 2), input_shape=(1, 28, 28)))
    cnn.add(LeakyReLU())
    cnn.add(Dropout(0.3))
    cnn.add(Convolution2D(64, 3, 3, border_mode='same', subsample=(1, 1)))
    cnn.add(LeakyReLU())
    cnn.add(Dropout(0.3))
    cnn.add(Convolution2D(128, 3, 3, border_mode='same', subsample=(1, 1)))
    cnn.add(LeakyReLU())
    cnn.add(Dropout(0.3))
    cnn.add(Convolution2D(256, 3, 3, border_mode='same', subsample=(1, 1)))
    cnn.add(LeakyReLU())
    cnn.add(Dropout(0.3))
    cnn.add(Flatten())
    image = Input(shape=(1, 28, 28))
    features = cnn(image)
    #有两个输出
    #输出真假值，范围在0~1
    fake = Dense(1, activation='sigmoid',name='generation')(features)
    #辅助分类器，输出图片分类
   aux = Dense(10, activation='softmax', name='auxiliary')(features)
    return Model(input=image, output=[fake, aux])

训练过程，50轮(epoch),把权重保存，每轮把虚假数据生成图处保存，观察虚假数据演化过程。

if __name__ =='__main__':
    #定义超参数
    nb_epochs = 50
    batch_size = 100
    latent_size = 100
    #优化器学习率
    adam_lr = 0.0002
    adam_beta_l = 0.5
    #构建判别网络
    discriminator = build_discriminator()
    discriminator.compile(optimizer=adam(lr=adam_lr, beta_l=adam_beta_l), loss='binary_crossentropy')
    latent = Input(shape=(lastent_size, ))
    image_class = Input(shape-(1, ), dtype='int32')
    #生成组合模型
    discriminator.trainable = False
    fake, aux = discriminator(fake)
    combined = Model(input=[latent, image_class], output=[fake, aux])
    combined.compile(optimizer=Adam(lr=adam_lr, beta_l=adam_beta_1), loss=['binary_crossentropy', 'sparse_categorical_crossentropy'])
    #将mnist数据转化为(...,1,28,28)维度,取值范围为[-1,1]
    (X_train,y_train),(X_test,y_test) = mnist.load_data()
    X_train = (X_train.astype(np.float32) - 127.5) / 127.5
    X_train = np.expand_dims(X_train, axis=1)
    X_test = (X_test.astype(np.float32) - 127.5) / 127.5
    X_test = np.expand_dims(X_test, axis=1)
    num_train, num_test = X_train.shape[0], X_test.shape[0]
    train_history = defaultdict(list)
    test_history = defaultdict(list)
    for epoch in range(epochs):
        print('Epoch {} of {}'.format(epoch + 1, epochs))
        num_batches = int(X_train.shape[0] / batch_size)
        progress_bar = Progbar(target=num_batches)
        epoch_gen_loss = []
        epoch_disc_loss = []
        for index in range(num_batches):
            progress_bar.update(index)
            #产生一个批次的噪声数据
            noise = np.random.uniform(-1, 1, (batch_size, latent_size))
            # 获取一个批次的真实数据
            image_batch = X_train[index * batch_size:(index + 1) * batch_size]
            label_batch = y_train[index * batch_size:(index + 1) * batch_size]
            # 生成一些噪声标记
            sampled_labels = np.random.randint(0, 10, batch_size)
            # 产生一个批次的虚假图片
            generated_images = generator.predict(
            [noise, sampled_labels.reshape((-1, 1))], verbose=0)
            X = np.concatenate((image_batch, generated_images))
            y = np.array([1] * batch_size + [0] * batch_size)
            aux_y = np.concatenate((label_batch, sampled_labels), axis=0)
            epoch_disc_loss.append(discriminator.train_on_batch(X, [y, aux_y]))
            # 产生两个批次噪声和标记
            noise = np.random.uniform(-1, 1, (2 * batch_size, latent_size))
            sampled_labels = np.random.randint(0, 10, 2 * batch_size)
            # 训练生成模型来欺骗判别模型，输出真/假都设为真
            trick = np.ones(2 * batch_size)
            epoch_gen_loss.append(combined.train_on_batch(
               [noise, sampled_labels.reshape((-1, 1))],
                [trick, sampled_labels]))
        print('\nTesting for epoch {}:'.format(epoch + 1))
        # 评估测试集，产生一个新批次噪声数据
        noise = np.random.uniform(-1, 1, (num_test, latent_size))
        sampled_labels = np.random.randint(0, 10, num_test)
        generated_images = generator.predict(
            [noise, sampled_labels.reshape((-1, 1))], verbose=False)
        X = np.concatenate((X_test, generated_images))
        y = np.array([1] * num_test + [0] * num_test)
        aux_y = np.concatenate((y_test, sampled_labels), axis=0)
        # 判别模型是否能判别
        discriminator_test_loss = discriminator.evaluate(
            X, [y, aux_y], verbose=False)
        discriminator_train_loss = np.mean(np.array(epoch_disc_loss), axis=0)
        # 创建两个批次新噪声数据
        noise = np.random.uniform(-1, 1, (2 * num_test, latent_size))
        sampled_labels = np.random.randint(0, 10, 2 * num_test)
        trick = np.ones(2 * num_test)
        generator_test_loss = combined.evaluate(
            [noise, sampled_labels.reshape((-1, 1))],
            [trick, sampled_labels], verbose=False)
        generator_train_loss = np.mean(np.array(epoch_gen_loss), axis=0)
        # 损失值等性能指标记录下来，并输出
        train_history['generator'].append(generator_train_loss)
        train_history['discriminator'].append(discriminator_train_loss)
        test_history['generator'].append(generator_test_loss)
        test_history['discriminator'].append(discriminator_test_loss)
        print('{0:<22s} | {1:4s} | {2:15s} | {3:5s}'.format(
            'component', *discriminator.metrics_names))
        print('-' * 65)
        ROW_FMT = '{0:<22s} | {1:<4.2f} | {2:<15.2f} | {3:<5.2f}'
        print(ROW_FMT.format('generator (train)',
                         *train_history['generator'][-1]))
        print(ROW_FMT.format('generator (test)',
                         *test_history['generator'][-1]))
        print(ROW_FMT.format('discriminator (train)',
                         *train_history['discriminator'][-1]))
        print(ROW_FMT.format('discriminator (test)',
                         *test_history['discriminator'][-1]))
        # 每个epoch保存一次权重
        generator.save_weights(
            'params_generator_epoch_{0:03d}.hdf5'.format(epoch), True)
        discriminator.save_weights(
            'params_discriminator_epoch_{0:03d}.hdf5'.format(epoch), True)
        # 生成一些可视化虚假数字看演化过程
        noise = np.random.uniform(-1, 1, (100, latent_size))
        sampled_labels = np.array([
            [i] * 10 for i in range(10)
        ]).reshape(-1, 1)
        generated_images = generator.predict(
            [noise, sampled_labels], verbose=0)
        # 整理到一个方格
        img = (np.concatenate([r.reshape(-1, 28)
                           for r in np.split(generated_images, 10)
                           ], axis=-1) * 127.5 + 127.5).astype(np.uint8)
        Image.fromarray(img).save(
            'plot_epoch_{0:03d}_generated.png'.format(epoch))
    pickle.dump({'train': train_history, 'test': test_history},
                open('acgan-history.pkl', 'wb'))

训练结束，创建3类文件。params_discriminator_epoch_{{epoch_number}}.hdf5,判别模型权重参数。params_generator_epoch_{{epoch_number}}.hdf5,生成模型权重参数。plot_epoch_{{epoch_number}}_generated.png 。

生成式对抗网络改进。生成式对抗网络(generative adversarial network,GAN)在无监督学习非常有效。常规生成式对抗网络判别器使用Sigmoid交叉熵损失函数，学习过程梯度消失。Wasserstein生成式对抗网络(Wasserstein generative adversarial network,WGAN)，使用Wasserstein距离度量，而不是Jensen-Shannon散度(Jensen-Shannon divergence,JSD)。使用最小二乘生成式对抗网络(least squares generative adversarial network,LSGAN)，判别模型用最小平方损失小函数(least squares loss function)。Sebastian Nowozin、Botond Cseke、Ryota Tomioka论文《f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization》。

参考资料：
《TensorFlow技术解析与实战》

欢迎付费咨询(150元每小时)，我的微信：qingxingfengzi