基於微軟開源深度學習演算法，用 Python 實現影象和視訊修復

model = networks.UNet(in_channels=1, out_channels=1, depth=4, conv_num=2, wf=6, padding=True, batch_norm=True, up_mode="upsample",with_tanh=False, sync_bn=True, antialiasing=True,)for image_name in imagelist:idx += 1print("processing", image_name)results = []scratch_image = Image.open(os.path.join(config.test_path, image_name)).convert("RGB")w, h = scratch_image.sizetransformed_image_PIL = data_transforms(scratch_image, config.input_size)scratch_image = transformed_image_PIL.convert("L")scratch_image = tv.transforms.ToTensor()(scratch_image)scratch_image = tv.transforms.Normalize([0.5], [0.5])(scratch_image)scratch_image = torch.unsqueeze(scratch_image, 0)scratch_image = scratch_image.to(config.GPU)P = torch.sigmoid(model(scratch_image))P = P.data.cpu()tv.utils.save_image((P >= 0.4).float(),os.path.join(output_dir, image_name[:-4] + ".png",),nrow=1,padding=0,normalize=True,)    transformed_image_PIL.save(os.path.join(input_dir, image_name[:-4] + ".png"))

def create_optimizers(self, opt):    G_params = list(self.netG.parameters())if opt.use_vae:        G_params += list(self.netE.parameters())if opt.isTrain:        D_params = list(self.netD.parameters())    beta1, beta2 = opt.beta1, opt.beta2if opt.no_TTUR:        G_lr, D_lr = opt.lr, opt.lrelse:        G_lr, D_lr = opt.lr / 2, opt.lr * 2    optimizer_G = torch.optim.Adam(G_params, lr=G_lr, betas=(beta1, beta2))    optimizer_D = torch.optim.Adam(D_params, lr=D_lr, betas=(beta1, beta2))return optimizer_G, optimizer_Ddef generate_fake(self, input_semantics, degraded_image, real_image, compute_kld_loss=False):    z = None    KLD_loss = Noneif self.opt.use_vae:        z, mu, logvar = self.encode_z(real_image)if compute_kld_loss:            KLD_loss = self.KLDLoss(mu, logvar) * self.opt.lambda_kld    fake_image = self.netG(input_semantics, degraded_image, z=z)    assert (not compute_kld_loss    ) or self.opt.use_vae, "You cannot compute KLD loss if opt.use_vae == False"return fake_image, KLD_lossdef discriminate(self, input_semantics, fake_image, real_image):if self.opt.no_parsing_map:        fake_concat = fake_image        real_concat = real_imageelse:        fake_concat = torch.cat([input_semantics, fake_image], dim=1)        real_concat = torch.cat([input_semantics, real_image], dim=1)    fake_and_real = torch.cat([fake_concat, real_concat], dim=0)    discriminator_out = self.netD(fake_and_real)    pred_fake, pred_real = self.divide_pred(discriminator_out)return pred_fake, pred_real