深度学习-第G5周：Pix2Pix详解

🍨 本文为[🔗365天深度学习训练营]内部限免文章（版权归 *K同学啊* 所有）
🍖 作者：[K同学啊]

本周任务：

1、了解并学习Pix2Pix算法

2、画出本文代码中Pix2Pix的生成器网络结构

前言

上篇文章介绍的CGAN是在GAN的基础上，在生成模型及判别模型中添加条件信息来引导模型的训练，实现生成内容可控。Pix2Pix则是在CGAN的基础上，用图像翻译的通用框架，旨在将一个图像的图像域中的图像转化成另一个图像中的图像域的图像。话比较绕，简而言之就是生成模型的输入跟判别模型的输入都是同一个图像+随机噪声z，

生成器

使用了U-NET的基础结构，实现压缩与扩张的连接，如下图：

##############################
#           U-NET
##############################


class UNetDown(nn.Module):
    def __init__(self, in_size, out_size, normalize=True, dropout=0.0):
        super(UNetDown, self).__init__()
        layers = [nn.Conv2d(in_size, out_size, 4, 2, 1, bias=False)]
        if normalize:
            layers.append(nn.InstanceNorm2d(out_size))
        layers.append(nn.LeakyReLU(0.2))
        if dropout:
            layers.append(nn.Dropout(dropout))
        self.model = nn.Sequential(*layers)

    def forward(self, x):
        return self.model(x)


class UNetUp(nn.Module):
    def __init__(self, in_size, out_size, dropout=0.0):
        super(UNetUp, self).__init__()
        layers = [
            nn.ConvTranspose2d(in_size, out_size, 4, 2, 1, bias=False),
            nn.InstanceNorm2d(out_size),
            nn.ReLU(inplace=True),
        ]
        if dropout:
            layers.append(nn.Dropout(dropout))

        self.model = nn.Sequential(*layers)

    def forward(self, x, skip_input):
        x = self.model(x)
        x = torch.cat((x, skip_input), 1)

        return x


class GeneratorUNet(nn.Module):
    def __init__(self, in_channels=3, out_channels=3):
        super(GeneratorUNet, self).__init__()

        self.down1 = UNetDown(in_channels, 64, normalize=False)
        self.down2 = UNetDown(64, 128)
        self.down3 = UNetDown(128, 256)
        self.down4 = UNetDown(256, 512, dropout=0.5)
        self.down5 = UNetDown(512, 512, dropout=0.5)
        self.down6 = UNetDown(512, 512, dropout=0.5)
        self.down7 = UNetDown(512, 512, dropout=0.5)
        self.down8 = UNetDown(512, 512, normalize=False, dropout=0.5)

        self.up1 = UNetUp(512, 512, dropout=0.5)
        self.up2 = UNetUp(1024, 512, dropout=0.5)
        self.up3 = UNetUp(1024, 512, dropout=0.5)
        self.up4 = UNetUp(1024, 512, dropout=0.5)
        self.up5 = UNetUp(1024, 256)
        self.up6 = UNetUp(512, 128)
        self.up7 = UNetUp(256, 64)

        self.final = 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):
        # U-Net generator with skip connections from encoder to decoder
        d1 = self.down1(x)
        d2 = self.down2(d1)
        d3 = self.down3(d2)
        d4 = self.down4(d3)
        d5 = self.down5(d4)
        d6 = self.down6(d5)
        d7 = self.down7(d6)
        d8 = self.down8(d7)
        u1 = self.up1(d8, d7)
        u2 = self.up2(u1, d6)
        u3 = self.up3(u2, d5)
        u4 = self.up4(u3, d4)
        u5 = self.up5(u4, d3)
        u6 = self.up6(u5, d2)
        u7 = self.up7(u6, d1)

        return self.final(u7)

判别器

##############################
#        Discriminator
##############################


class Discriminator(nn.Module):
    def __init__(self, in_channels=3):
        super(Discriminator, self).__init__()

        def discriminator_block(in_filters, out_filters, normalization=True):
            """Returns downsampling layers of each discriminator block"""
            layers = [nn.Conv2d(in_filters, out_filters, 4, stride=2, padding=1)]
            if normalization:
                layers.append(nn.InstanceNorm2d(out_filters))
            layers.append(nn.LeakyReLU(0.2, inplace=True))
            return layers

        self.model = nn.Sequential(
            *discriminator_block(in_channels * 2, 64, normalization=False),
            *discriminator_block(64, 128),
            *discriminator_block(128, 256),
            *discriminator_block(256, 512),
            nn.ZeroPad2d((1, 0, 1, 0)),
            nn.Conv2d(512, 1, 4, padding=1, bias=False)
        )

    def forward(self, img_A, img_B):
        # Concatenate image and condition image by channels to produce input
        img_input = torch.cat((img_A, img_B), 1)
        return self.model(img_input)

在 Discriminator中，Pix2Pix算法提出将图像分成N X N个图像块，分别计算图像纹理的损失，我们可以通过调节不同大小的N值来得到最适合的模型感受野

超参数及数据准备

import argparse
import time
import datetime
import sys

import torchvision.transforms as transforms
from torchvision.utils import save_image
from torch.utils.data import DataLoader
from torch.autograd import Variable

from models import *
from datasets import *

import torch.nn as nn
import torch.nn.functional as F
import torch

parser = argparse.ArgumentParser()
parser.add_argument("--epoch", type=int, default=0, help="epoch to start training from")
parser.add_argument("--n_epochs", type=int, default=100, help="number of epochs of training")
parser.add_argument("--dataset_name", type=str, default="data_facades", help="name of the dataset")
parser.add_argument("--batch_size", type=int, default=1, help="size of the batches")
parser.add_argument("--lr", type=float, default=0.0002, help="adam: learning rate")
parser.add_argument("--b1", type=float, default=0.5, help="adam: decay of first order momentum of gradient")
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
parser.add_argument("--decay_epoch", type=int, default=100, help="epoch from which to start lr decay")
parser.add_argument("--n_cpu", type=int, default=8, help="number of cpu threads to use during batch generation")
parser.add_argument("--img_height", type=int, default=256, help="size of image height")
parser.add_argument("--img_width", type=int, default=256, help="size of image width")
parser.add_argument("--channels", type=int, default=3, help="number of image channels")
parser.add_argument(
    "--sample_interval", type=int, default=500, help="interval between sampling of images from generators"
)
parser.add_argument("--checkpoint_interval", type=int, default=-1, help="interval between model checkpoints")
opt = parser.parse_args()
print(opt)

os.makedirs("images/%s" % opt.dataset_name, exist_ok=True)
os.makedirs("saved_models/%s" % opt.dataset_name, exist_ok=True)

cuda = True if torch.cuda.is_available() else False

# Loss functions
criterion_GAN = torch.nn.MSELoss()
criterion_pixelwise = torch.nn.L1Loss()

# Loss weight of L1 pixel-wise loss between translated image and real image
lambda_pixel = 100

# Calculate output of image discriminator (PatchGAN)
patch = (1, opt.img_height // 2 ** 4, opt.img_width // 2 ** 4)

# Initialize generator and discriminator
generator = GeneratorUNet()
discriminator = Discriminator()

if cuda:
    generator = generator.cuda()
    discriminator = discriminator.cuda()
    criterion_GAN.cuda()
    criterion_pixelwise.cuda()

if opt.epoch != 0:
    # Load pretrained models
    generator.load_state_dict(torch.load("saved_models/%s/generator_%d.pth" % (opt.dataset_name, opt.epoch)))
    discriminator.load_state_dict(torch.load("saved_models/%s/discriminator_%d.pth" % (opt.dataset_name, opt.epoch)))
else:
    # Initialize weights
    generator.apply(weights_init_normal)
    discriminator.apply(weights_init_normal)

# Optimizers
optimizer_G = torch.optim.Adam(generator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))

# Configure dataloaders
transforms_ = [
    transforms.Resize((opt.img_height, opt.img_width), Image.BICUBIC),
    transforms.ToTensor(),
    transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]

dataloader = DataLoader(
    ImageDataset("./%s" % opt.dataset_name, transforms_=transforms_),
    batch_size=opt.batch_size,
    shuffle=True,
    num_workers=opt.n_cpu,
)

val_dataloader = DataLoader(
    ImageDataset("./%s" % opt.dataset_name, transforms_=transforms_, mode="val"),
    batch_size=10,
    shuffle=True,
    num_workers=1,
)

# Tensor type
Tensor = torch.cuda.FloatTensor if cuda else torch.FloatTensor


def sample_images(batches_done):
    """Saves a generated sample from the validation set"""
    imgs = next(iter(val_dataloader))
    real_A = Variable(imgs["B"].type(Tensor))
    real_B = Variable(imgs["A"].type(Tensor))
    fake_B = generator(real_A)
    img_sample = torch.cat((real_A.data, fake_B.data, real_B.data), -2)
    save_image(img_sample, "images/%s/%s.png" % (opt.dataset_name, batches_done), nrow=5, normalize=True)

训练代码

# ----------
#  Training
# ----------

if __name__ == '__main__':
    prev_time = time.time()

    for epoch in range(opt.epoch, opt.n_epochs):
        for i, batch in enumerate(dataloader):

            # Model inputs
            real_A = Variable(batch["B"].type(Tensor))
            real_B = Variable(batch["A"].type(Tensor))

            # Adversarial ground truths
            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)
            # Pixel-wise loss
            loss_pixel = criterion_pixelwise(fake_B, real_B)

            # Total loss
            loss_G = loss_GAN + lambda_pixel * loss_pixel

            loss_G.backward()

            optimizer_G.step()

            # ---------------------
            #  Train Discriminator
            # ---------------------

            optimizer_D.zero_grad()

            # Real loss
            pred_real = discriminator(real_B, real_A)
            loss_real = criterion_GAN(pred_real, valid)

            # Fake loss
            pred_fake = discriminator(fake_B.detach(), real_A)
            loss_fake = criterion_GAN(pred_fake, fake)

            # Total loss
            loss_D = 0.5 * (loss_real + loss_fake)

            loss_D.backward()
            optimizer_D.step()

            # --------------
            #  Log Progress
            # --------------

            # Determine approximate time left
            batches_done = epoch * len(dataloader) + i
            batches_left = opt.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,
                    opt.n_epochs,
                    i,
                    len(dataloader),
                    loss_D.item(),
                    loss_G.item(),
                    loss_pixel.item(),
                    loss_GAN.item(),
                    time_left,
                )
            )

            # If at sample interval save image
            if batches_done % opt.sample_interval == 0:
                sample_images(batches_done)

        if opt.checkpoint_interval != -1 and epoch % opt.checkpoint_interval == 0:
            # Save model checkpoints
            torch.save(generator.state_dict(), "saved_models/%s/generator_%d.pth" % (opt.dataset_name, epoch))
            torch.save(discriminator.state_dict(), "saved_models/%s/discriminator_%d.pth" % (opt.dataset_name, epoch))

为保存模型，将超参数中parser.add_argument("--checkpoint_interval", type=int, default=5, help="interval between model checkpoints")，-1改为5，每个epoch 保存一次模型

运行结果

0.png

15000.png

33000.png

50500.png