【源码解读】pix2pix（一）：训练

源码地址：https://github.com/mrzhu-cool/pix2pix-pytorch

相比于朱俊彦的版本，这一版更加简单易读

训练的代码在train.py，开头依然是很多代码的共同三板斧，加载参数，加载数据，加载模型

命令行参数

# Training settings
parser = argparse.ArgumentParser(description='pix2pix-pytorch-implementation')
parser.add_argument('--dataset', required=True, help='facades')
parser.add_argument('--batch_size', type=int, default=1, help='training batch size')
parser.add_argument('--test_batch_size', type=int, default=1, help='testing batch size')
parser.add_argument('--direction', type=str, default='b2a', help='a2b or b2a')
parser.add_argument('--input_nc', type=int, default=3, help='input image channels')
parser.add_argument('--output_nc', type=int, default=3, help='output image channels')
parser.add_argument('--ngf', type=int, default=64, help='generator filters in first conv layer')
parser.add_argument('--ndf', type=int, default=64, help='discriminator filters in first conv layer')
parser.add_argument('--epoch_count', type=int, default=1, help='the starting epoch count')
parser.add_argument('--niter', type=int, default=100, help='# of iter at starting learning rate')
parser.add_argument('--niter_decay', type=int, default=100, help='# of iter to linearly decay learning rate to zero')
parser.add_argument('--lr', type=float, default=0.0002, help='initial learning rate for adam')
parser.add_argument('--lr_policy', type=str, default='lambda', help='learning rate policy: lambda|step|plateau|cosine')
parser.add_argument('--lr_decay_iters', type=int, default=50, help='multiply by a gamma every lr_decay_iters iterations')
parser.add_argument('--beta1', type=float, default=0.5, help='beta1 for adam. default=0.5')
parser.add_argument('--cuda', action='store_true', help='use cuda?')
parser.add_argument('--threads', type=int, default=4, help='number of threads for data loader to use')
parser.add_argument('--seed', type=int, default=123, help='random seed to use. Default=123')
parser.add_argument('--lamb', type=int, default=10, help='weight on L1 term in objective')
opt = parser.parse_args()

数据

print('===> Loading datasets')
root_path = "dataset/"
train_set = get_training_set(root_path + opt.dataset, opt.direction)
test_set = get_test_set(root_path + opt.dataset, opt.direction)
training_data_loader = DataLoader(dataset=train_set, num_workers=opt.threads, batch_size=opt.batch_size, shuffle=True)
testing_data_loader = DataLoader(dataset=test_set, num_workers=opt.threads, batch_size=opt.test_batch_size, shuffle=False)

模型

print('===> Building models')
net_g = define_G(opt.input_nc, opt.output_nc, opt.ngf, 'batch', False, 'normal', 0.02, gpu_id=device)
net_d = define_D(opt.input_nc + opt.output_nc, opt.ndf, 'basic', gpu_id=device)

优化器，损失函数

criterionGAN = GANLoss().to(device)
criterionL1 = nn.L1Loss().to(device)
criterionMSE = nn.MSELoss().to(device)

# setup optimizer
optimizer_g = optim.Adam(net_g.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
optimizer_d = optim.Adam(net_d.parameters(), lr=opt.lr, betas=(opt.beta1, 0.999))
net_g_scheduler = get_scheduler(optimizer_g, opt)
net_d_scheduler = get_scheduler(optimizer_d, opt)

接着按批次读取数据，首先更新判别器，判别器的输入是图像对（真，真）（真，假）

######################
        # (1) Update D network
        ######################

        optimizer_d.zero_grad()

        # train with fake
        fake_ab = torch.cat((real_a, fake_b), 1)
        pred_fake = net_d.forward(fake_ab.detach())
        loss_d_fake = criterionGAN(pred_fake, False)

        # train with real
        real_ab = torch.cat((real_a, real_b), 1)
        pred_real = net_d.forward(real_ab)
        loss_d_real = criterionGAN(pred_real, True)

        # Combined D loss
        loss_d = (loss_d_fake + loss_d_real) * 0.5

        loss_d.backward()

        optimizer_d.step()

然后更新生成器，生成器的损失由判别器产生的损失函数和真假图像之间的L1约束组成

######################
        # (2) Update G network
        ######################

        optimizer_g.zero_grad()

        # First, G(A) should fake the discriminator
        fake_ab = torch.cat((real_a, fake_b), 1)
        pred_fake = net_d.forward(fake_ab)
        loss_g_gan = criterionGAN(pred_fake, True)

        # Second, G(A) = B
        loss_g_l1 = criterionL1(fake_b, real_b) * opt.lamb

        loss_g = loss_g_gan + loss_g_l1

        loss_g.backward()

        optimizer_g.step()

最后更新学习率

update_learning_rate(net_g_scheduler, optimizer_g)
update_learning_rate(net_d_scheduler, optimizer_d)

比较核心的代码是网络构造，以及一些工具函数，放在后面写