生成式對(duì)抗網(wǎng)絡(luò)(gennerative adversarial network,GAN)吻贿，谷歌2014年提出網(wǎng)絡(luò)模型串结。靈感自二人博弈的零和博弈，目前最火的非監(jiān)督深度學(xué)習(xí)舅列。GAN之父肌割，Ian J.Goodfellow，公認(rèn)人工智能頂級(jí)專家帐要。
原理把敞。
生成式對(duì)搞網(wǎng)絡(luò)包含一個(gè)生成模型(generative model,G)和一個(gè)判別模型(discriminative model,D)。Ian J.Goodfellow榨惠、Jean Pouget-Abadie奋早、Mehdi Mirza、Bing Xu赠橙、David Warde-Farley耽装、Sherjil Ozair、Aaron Courville期揪、Yoshua Bengio論文掉奄，《Generative Adversarial Network》，https://arxiv.org/abs/1406.2661 横侦。
生成式對(duì)抗網(wǎng)絡(luò)結(jié)構(gòu)：
噪聲數(shù)據(jù)->生成模型->假圖片---|
|->判別模型->真/假
打亂訓(xùn)練數(shù)據(jù)->訓(xùn)練集->真圖片-|
生成式對(duì)抗網(wǎng)絡(luò)主要解決如何從訓(xùn)練樣本中學(xué)習(xí)出新樣本挥萌。生成模型負(fù)責(zé)訓(xùn)練出樣本的分布，如果訓(xùn)練樣本是圖片就生成相似的圖片枉侧，如果訓(xùn)練樣本是文章名子就生成相似的文章名子引瀑。判別模型是一個(gè)二分類器，用來判斷輸入樣本是真實(shí)數(shù)據(jù)還是訓(xùn)練生成的樣本榨馁。
生成式對(duì)抗網(wǎng)絡(luò)優(yōu)化憨栽，是一個(gè)二元極小極大博弈(minimax two-player game)問題。使生成模型輸出在輸入給判別模型時(shí)，判斷模型秀難判斷是真實(shí)數(shù)據(jù)還是虛似數(shù)據(jù)屑柔。訓(xùn)練好的生成模型屡萤，能把一個(gè)噪聲向量轉(zhuǎn)化成和訓(xùn)練集類似的樣本。Argustus Odena掸宛、Christopher Olah死陆、Jonathon Shlens論文《Coditional Image Synthesis with Auxiliary Classifier GANs》。
輔助分類器生成式對(duì)抗網(wǎng)絡(luò)(auxiliary classifier GAN唧瘾，AC-GAN)實(shí)現(xiàn)措译。
生成式對(duì)抗網(wǎng)絡(luò)應(yīng)用。生成數(shù)字饰序，生成人臉圖像领虹。
生成式對(duì)抗網(wǎng)絡(luò)實(shí)現(xiàn)。https://github.com/fchollet/keras/blob/master/examples/mnist_acgan.py 求豫。
Augustus Odena塌衰、Chistopher Olah和Jonathon Shlens 論文《Conditional Image Synthesis With Auxiliary Classifier GANs》。
通過噪聲蝠嘉，讓生成模型G生成虛假數(shù)據(jù)最疆，和真實(shí)數(shù)據(jù)一起送到判別模型D，判別模型一方面輸出數(shù)據(jù)真/假,一方面輸出圖片分類是晨。
首先定義生成模型肚菠，目的是生成一對(duì)(z,L)數(shù)據(jù)舔箭，z是噪聲向量罩缴，L是(1,28,28)的圖像空間。
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)))
#上采樣层扶，圖你尺寸變?yōu)?14X14
cnn.add(UpSampling2D(size=(2,2)))
cnn.add(Convolution2D(256, 5, 5, border_mode='same', activation='relu', init='glorot_normal'))
#上采樣箫章，圖像尺寸變?yōu)?8X28
cnn.add(UpSampling2D(size=(2,2)))
cnn.add(Convolution2D(128, 5, 5, border_mode='same', activation='relu', init='glorot_normal'))
#規(guī)約到1個(gè)通道
cnn.add(Convolution2D(1, 2, 2, border_mode='same', activation='tanh', init='glorot_normal'))
#生成模型輸入層,特征向量
latent = Input(shape=(latent_size, ))
#生成模型輸入層，標(biāo)記
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) #輸出虛假圖片
return Model(input=[latent, image_class], output=fake_image)
定義判別模型镜会，輸入(1,28,28)圖片檬寂，輸出兩個(gè)值，一個(gè)是判別模型認(rèn)為這張圖片是否是虛假圖片戳表，另一個(gè)是判別模型認(rèn)為這第圖片所屬分類桶至。
def build_discriminator();
#采用激活函數(shù)Leaky ReLU來替換標(biāo)準(zhǔn)的卷積神經(jīng)網(wǎng)絡(luò)中的激活函數(shù)
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)
#有兩個(gè)輸出
#輸出真假值，范圍在0~1
fake = Dense(1, activation='sigmoid',name='generation')(features)
#輔助分類器匾旭，輸出圖片分類
aux = Dense(10, activation='softmax', name='auxiliary')(features)
return Model(input=image, output=[fake, aux])
訓(xùn)練過程镣屹，50輪(epoch),把權(quán)重保存，每輪把虛假數(shù)據(jù)生成圖處保存价涝，觀察虛假數(shù)據(jù)演化過程女蜈。
if __name__ =='__main__':
#定義超參數(shù)
nb_epochs = 50
batch_size = 100
latent_size = 100
#優(yōu)化器學(xué)習(xí)率
adam_lr = 0.0002
adam_beta_l = 0.5
#構(gòu)建判別網(wǎng)絡(luò)
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')
#生成組合模型
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'])
#將mnist數(shù)據(jù)轉(zhuǎn)化為(...,1,28,28)維度,取值范圍為[-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)
#產(chǎn)生一個(gè)批次的噪聲數(shù)據(jù)
noise = np.random.uniform(-1, 1, (batch_size, latent_size))
# 獲取一個(gè)批次的真實(shí)數(shù)據(jù)
image_batch = X_train[index * batch_size:(index + 1) * batch_size]
label_batch = y_train[index * batch_size:(index + 1) * batch_size]
# 生成一些噪聲標(biāo)記
sampled_labels = np.random.randint(0, 10, batch_size)
# 產(chǎn)生一個(gè)批次的虛假圖片
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]))
# 產(chǎn)生兩個(gè)批次噪聲和標(biāo)記
noise = np.random.uniform(-1, 1, (2 * batch_size, latent_size))
sampled_labels = np.random.randint(0, 10, 2 * batch_size)
# 訓(xùn)練生成模型來欺騙判別模型，輸出真/假都設(shè)為真
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))
# 評(píng)估測試集，產(chǎn)生一個(gè)新批次噪聲數(shù)據(jù)
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)
# 判別模型是否能判別
discriminator_test_loss = discriminator.evaluate(
X, [y, aux_y], verbose=False)
discriminator_train_loss = np.mean(np.array(epoch_disc_loss), axis=0)
# 創(chuàng)建兩個(gè)批次新噪聲數(shù)據(jù)
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)
# 損失值等性能指標(biāo)記錄下來伪窖，并輸出
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]))
# 每個(gè)epoch保存一次權(quán)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)
# 生成一些可視化虛假數(shù)字看演化過程
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)
# 整理到一個(gè)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'))
訓(xùn)練結(jié)束逸寓，創(chuàng)建3類文件。params_discriminator_epoch_{{epoch_number}}.hdf5,判別模型權(quán)重參數(shù)覆山。params_generator_epoch_{{epoch_number}}.hdf5,生成模型權(quán)重參數(shù)竹伸。plot_epoch_{{epoch_number}}_generated.png 。
生成式對(duì)抗網(wǎng)絡(luò)改進(jìn)簇宽。生成式對(duì)抗網(wǎng)絡(luò)(generative adversarial network,GAN)在無監(jiān)督學(xué)習(xí)非常有效佩伤。常規(guī)生成式對(duì)抗網(wǎng)絡(luò)判別器使用Sigmoid交叉熵?fù)p失函數(shù)，學(xué)習(xí)過程梯度消失晦毙。Wasserstein生成式對(duì)抗網(wǎng)絡(luò)(Wasserstein generative adversarial network,WGAN)生巡，使用Wasserstein距離度量，而不是Jensen-Shannon散度(Jensen-Shannon divergence,JSD)见妒。使用最小二乘生成式對(duì)抗網(wǎng)絡(luò)(least squares generative adversarial network,LSGAN)孤荣，判別模型用最小平方損失小函數(shù)(least squares loss function)。Sebastian Nowozin须揣、Botond Cseke盐股、Ryota Tomioka論文《f-GAN: Training Generative Neural Samplers using Variational Divergence Minimization》。
參考資料：
《TensorFlow技術(shù)解析與實(shí)戰(zhàn)》
歡迎付費(fèi)咨詢(150元每小時(shí))耻卡，我的微信：qingxingfengzi
