比照两个框架实现同一个模型到底有什么区别？
第一步，咱们对数据集进行图像预处理。咱们在这里抉择 Facades 数据集，咱们将 2 张图像合并为一张，以便在训练过程中进行一些加强。
Pytorch：
def __getitem__(self, index):img = Image.open(self.files[index % len(self.files)]) w, h = img.size img_A = img.crop((0, 0, w / 2, h)) img_B = img.crop((w / 2, 0, w, h))if np.random.random() < 0.5: img_A = Image.fromarray(np.array(img_A)[:, ::-1, :], "RGB") img_B = Image.fromarray(np.array(img_B)[:, ::-1, :], "RGB")img_A = self.transform(img_A) img_B = self.transform(img_B)return {"A": img_A, "B": img_B}Keras：
def load_batch(self, batch_size=1, is_testing=False): data_type = "train" if not is_testing else "val" path = glob('./datasets/%s/%s/*' % (self.dataset_name, data_type))self.n_batches = int(len(path) / batch_size)for i in range(self.n_batches-1): batch = path[i*batch_size:(i+1)*batch_size] imgs_A, imgs_B = [], [] for img in batch: img = self.imread(img) h, w, _ = img.shape half_w = int(w/2) img_A = img[:, :half_w, :] img_B = img[:, half_w:, :]img_A = resize(img_A, self.img_res) img_B = resize(img_B, self.img_res)if not is_testing and np.random.random() > 0.5: img_A = np.fliplr(img_A) img_B = np.fliplr(img_B)imgs_A.append(img_A) imgs_B.append(img_B)imgs_A = np.array(imgs_A)/127.5 - 1. imgs_B = np.array(imgs_B)/127.5 - 1.yield imgs_A, imgs_B模型
在论文中提到应用的模型是 U-Net，所以须要应用层间的跳跃连贯（恒等函数）。应用上采样和下采样卷积制作自编码器生成和判断模型。
Pytorch：
class Generator(nn.Module): def __init__(self, in_channels=3, out_channels=3): super(Generator, self).__init__() self.down1 = DownSampleConv(in_channels, 64, batchnorm=False) self.down2 = DownSampleConv(64, 128) self.down3 = DownSampleConv(128, 256) self.down4 = DownSampleConv(256, 512, dropout_rate=0.5) self.down5 = DownSampleConv(512, 512, dropout_rate=0.5) self.down6 = DownSampleConv(512, 512, dropout_rate=0.5) self.down7 = DownSampleConv(512, 512, dropout_rate=0.5) self.down8 = DownSampleConv(512, 512, batchnorm=False, dropout_rate=0.5)self.up1 = UpSampleConv(512, 512, dropout_rate=0.5) self.up2 = UpSampleConv(1024, 512, dropout_rate=0.5) self.up3 = UpSampleConv(1024, 512, dropout_rate=0.5) self.up4 = UpSampleConv(1024, 512, dropout_rate=0.5) self.up5 = UpSampleConv(1024, 256) self.up6 = UpSampleConv(512, 128) self.up7 = UpSampleConv(256, 64)self.last_conv = nn.Sequential( nn.Upsample(scale_factor=2), nn.ZeroPad2d((1, 0, 1, 0)), nn.Conv2d(128, out_channels, 4, padding=1), nn.Tanh(), )def forward(self, x): ds1 = self.down1(x) ds2 = self.down2(ds1) ds3 = self.down3(ds2) ds4 = self.down4(ds3) ds5 = self.down5(ds4) ds6 = self.down6(ds5) ds7 = self.down7(ds6) ds8 = self.down8(ds7) us1 = self.up1(ds8, ds7) us2 = self.up2(us1, ds6) us3 = self.up3(us2, ds5) us4 = self.up4(us3, ds4) us5 = self.up5(us4, ds3) us6 = self.up6(us5, ds2) us7 = self.up7(us6, ds1) return self.last_conv(us7)class Discriminator(nn.Module): def __init__(self, in_channels=3): super(Discriminator, self).__init__()self.model = nn.Sequential( DownSampleConv(in_channels + in_channels, 64, batchnorm=False, inplace=True), DownSampleConv(64, 128, inplace=True), DownSampleConv(128, 256, inplace=True), DownSampleConv(256, 512, inplace=True), nn.ZeroPad2d((1, 0, 1, 0)), nn.Conv2d(512, 1, (4, 4), padding=1, bias=False) )def forward(self, x, y): img_input = torch.cat([x, y], 1) return self.model(img_input)Keras
def build_generator(self): initializers = RandomNormal(stddev=0.02) input_image = Input(shape=self.img_shape) e1 = self.encoder_block(input_image, 64, batchnorm=False) e2 = self.encoder_block(e1, 128) e3 = self.encoder_block(e2, 256) e4 = self.encoder_block(e3, 512) e5 = self.encoder_block(e4, 512) e6 = self.encoder_block(e5, 512) e7 = self.encoder_block(e6, 512)d2 = self.decoder_block(e7, e6, 512) d3 = self.decoder_block(d2, e5, 512) d4 = self.decoder_block(d3, e4, 512) d5 = self.decoder_block(d4, e3, 256) d6 = self.decoder_block(d5, e2, 128) d7 = self.decoder_block(d6, e1, 64) up = UpSampling2D(size=2)(d7) output_image = Conv2D(self.channels, (4, 4), strides=1, padding='same', kernel_initializer=initializers, activation='tanh')(up) model = Model(input_image, output_image) return modeldef build_discriminator(self): initializers = RandomNormal(stddev=0.02) input_source_image = Input(self.img_shape) input_target_image = Input(self.img_shape) merged_input = Concatenate(axis=-1)([input_source_image, input_target_image]) filters_list = [64, 128, 256, 512]def disc_layer(input_layer, filters, kernel_size=(4, 4), batchnorm=True): x = Conv2D(filters, kernel_size=kernel_size, strides=2, padding='same', kernel_initializer=initializers)( input_layer) x = LeakyReLU(0.2)(x) if batchnorm: x = BatchNormalization()(x) return xx = disc_layer(merged_input, filters_list[0], batchnorm=False) x = disc_layer(x, filters_list[1]) x = disc_layer(x, filters_list[2]) x = disc_layer(x, filters_list[3])discriminator_output = Conv2D(1, kernel_size=(4, 4), padding='same', kernel_initializer=initializers)(x) model = Model([input_source_image, input_target_image], discriminator_output) return model训练过程
对于训练，咱们持续应用生成器和鉴别器架构。应用论文中倡议的权重初始化办法更改权重初始化器（权重从均值为 0 的高斯分布初始化， 标准差 0.02)。此外还有一些训练的超参数。(Adam优化器，LR=0.0002, B1=0.5, B2=0.999)
Pytorch：
# Training prev_time = time.time() for epoch in range(init_epoch, n_epochs): for i, batch in enumerate(dataloader):real_A = Variable(batch["B"].type(Tensor)) real_B = Variable(batch["A"].type(Tensor))valid = Variable(Tensor(np.ones((real_A.size(0), *patch))), requires_grad=False) fake = Variable(Tensor(np.zeros((real_A.size(0), *patch))), requires_grad=False)# Train Generators optimizer_G.zero_grad()# GAN loss fake_B = generator(real_A) pred_fake = discriminator(fake_B, real_A) loss_GAN = criterion_GAN(pred_fake, valid) loss_pixel = criterion_pixelwise(fake_B, real_B) loss_G = loss_GAN + lambda_pixel * loss_pixel loss_G.backward() optimizer_G.step()# Train Discriminator optimizer_D.zero_grad() pred_real = discriminator(real_B, real_A) loss_real = criterion_GAN(pred_real, valid) pred_fake = discriminator(fake_B.detach(), real_A) loss_fake = criterion_GAN(pred_fake, fake) loss_D = (loss_real + loss_fake) * 0.5 loss_D.backward() optimizer_D.step()batches_done = epoch * len(dataloader) + i batches_left = n_epochs * len(dataloader) - batches_done time_left = datetime.timedelta(seconds=batches_left * (time.time() - prev_time)) prev_time = time.time()# Print log sys.stdout.write( "\r[Epoch %d/%d] [Batch %d/%d] [D loss: %f] [G loss: %f, pixel: %f, adv: %f] ETA: %s" % (epoch, n_epochs, i, len(dataloader), loss_D.item(), loss_G.item(), loss_pixel.item(), loss_GAN.item(), time_left) ) G_losses.append(loss_G.item()) D_losses.append(loss_D.item())Keras：
def train(self, epochs, batch_size=1, sample_interval=50): valid = np.ones((batch_size,) + self.disc_patch) fake = np.zeros((batch_size,) + self.disc_patch)for epoch in range(epochs): for batch_i, (imgs_A, imgs_B) in enumerate(self.data_loader.load_batch(batch_size)):fake_A = self.generator.predict(imgs_B)d_loss_real = self.discriminator.train_on_batch([imgs_A, imgs_B], valid) d_loss_fake = self.discriminator.train_on_batch([fake_A, imgs_B], fake) d_loss = 0.5 * np.add(d_loss_real, d_loss_fake)g_loss = self.combined.train_on_batch([imgs_A, imgs_B], [valid, imgs_A])print( "[Epoch %d/%d] [Batch %d/%d] [D loss: %f,D Loss Real: %f, D Loss Fake: %f, acc: %3d%%] [G loss: %f, L1 Loss: %f, acc: %3d%%]" % ( epoch, epochs, batch_i, self.data_loader.n_batches, d_loss[0], d_loss_real[0], d_loss_fake[0], 100 * d_loss[1], g_loss[0], g_loss[1], 100 * g_loss[2])) self.G_losses.append(g_loss[0]) self.D_losses.append(d_loss[0]) if batch_i % sample_interval == 0: self.sample_images(epoch, batch_i) self.plot_metrics(self.G_losses, self.D_losses)这样咱们就实现了一个残缺的流程，能够看到，除了训练这块其余步骤两个框架的差异根本不大。如果你对残缺代码感兴趣，本文代码如下：
https://avoid.overfit.cn/post/be317cf0a41c4a48b8a80398489120b3
作者:vargha khallokhi