摘要

Despite the breakthroughs in accuracy and speed of single image super-resolution using faster and deeper convolutional neural networks, one central problem remains largely unsolved: how do we recover the finer texture details when we super-resolve at large upscaling factors? The behavior of optimization-based super-resolution methods is
principally driven by the choice of the objective function.Recent work has largely focused on minimizing the mean squared reconstruction error. The resulting estimates have
high peak signal-to-noise ratios, but they are often lacking high-frequency details and are perceptually unsatisfying in the sense that they fail to match the fidelity expected at
the higher resolution. In this paper, we present SRGAN,a generative adversarial network (GAN) for image superresolution (SR). To our knowledge, it is the first framework capable of inferring photo-realistic natural images for 4×upscaling factors. To achieve this, we propose a perceptual loss function which consists of an adversarial loss and a content loss. The adversarial loss pushes our solution to the natural image manifold using a discriminator network that is trained to differentiate between the super-resolved
images and original photo-realistic images. In addition, we use a content loss motivated by perceptual similarity instead of similarity in pixel space. Our deep residual network
is able to recover photo-realistic textures from heavily downsampled images on public benchmarks. An extensive mean-opinion-score (MOS) test shows hugely significant
gains in perceptual quality using SRGAN. The MOS scores obtained with SRGAN are closer to those of the original high-resolution images than to those obtained with any
state-of-the-art method.

论文主要讲述GAN应用于超分辨率的研究。作者们命名了一种对抗神经网络为SRGAN，它可以很好地将低分辨率图像恢复成原始高分辨率图像。在MOS测试中，SRGAN取得了比传统方法都更高的得分。

论文主体

https://arxiv.org/pdf/1609.04802v5.pdf

实现代码

读这篇论文的主要工作就是把论文中的方法复现一下，github中有许多同类型的开源项目，但由于版本更迭的问题，直接down下来的代码往往会有许多bug。在这里记录一下调试代码的全过程。

Dataset

数据集下载网址：http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
celebA是一个人脸数据集，包含几十万张不同的人脸图片，同时这些数据已经是裁剪好的，并且经过压缩后占用空间都非常小，非常适合进行需要大量人脸图片的训练。

之后构建dataset.py来读取dataset。

dataset.py

import glob
import random
import os
import numpy as npimport torch
from torch.utils.data import Dataset
from PIL import Image
import torchvision.transforms as transforms# Normalization parameters for pre-trained PyTorch models
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])class ImageDataset(Dataset):def __init__(self, root, hr_shape):hr_height, hr_width = hr_shape# Transforms for low resolution images and high resolution imagesself.lr_transform = transforms.Compose([transforms.Resize((hr_height // 4, hr_height // 4), Image.BICUBIC),transforms.ToTensor(),transforms.Normalize(mean, std),])self.hr_transform = transforms.Compose([transforms.Resize((hr_height, hr_height), Image.BICUBIC),transforms.ToTensor(),transforms.Normalize(mean, std),])self.files = sorted(glob.glob(root + "/*.*"))def __getitem__(self, index):img = Image.open(self.files[index % len(self.files)])img_lr = self.lr_transform(img)img_hr = self.hr_transform(img)return {"lr": img_lr, "hr": img_hr}def __len__(self):return len(self.files)

注意这里指定的读取路径是文件中的每一个图片，也就是如果用data/img_align_celeba文件夹来存储数据的话，需要将图片全部解压放置在data/img_align_celeba中。

Model

直接用github中写好的论文中的网络结构

models.py

import torch.nn as nn
import torch.nn.functional as F
import torch
from torchvision.models import vgg19
import mathclass FeatureExtractor(nn.Module):def __init__(self):super(FeatureExtractor, self).__init__()vgg19_model = vgg19(pretrained=True)self.feature_extractor = nn.Sequential(*list(vgg19_model.features.children())[:18])def forward(self, img):return self.feature_extractor(img)class ResidualBlock(nn.Module):def __init__(self, in_features):super(ResidualBlock, self).__init__()self.conv_block = nn.Sequential(nn.Conv2d(in_features, in_features, kernel_size=3, stride=1, padding=1),nn.BatchNorm2d(in_features, 0.8),nn.PReLU(),nn.Conv2d(in_features, in_features, kernel_size=3, stride=1, padding=1),nn.BatchNorm2d(in_features, 0.8),)def forward(self, x):return x + self.conv_block(x)class GeneratorResNet(nn.Module):def __init__(self, in_channels=3, out_channels=3, n_residual_blocks=16):super(GeneratorResNet, self).__init__()# First layerself.conv1 = nn.Sequential(nn.Conv2d(in_channels, 64, kernel_size=9, stride=1, padding=4), nn.PReLU())# Residual blocksres_blocks = []for _ in range(n_residual_blocks):res_blocks.append(ResidualBlock(64))self.res_blocks = nn.Sequential(*res_blocks)# Second conv layer post residual blocksself.conv2 = nn.Sequential(nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1), nn.BatchNorm2d(64, 0.8))# Upsampling layersupsampling = []for out_features in range(2):upsampling += [# nn.Upsample(scale_factor=2),nn.Conv2d(64, 256, 3, 1, 1),nn.BatchNorm2d(256),nn.PixelShuffle(upscale_factor=2),nn.PReLU(),]self.upsampling = nn.Sequential(*upsampling)# Final output layerself.conv3 = nn.Sequential(nn.Conv2d(64, out_channels, kernel_size=9, stride=1, padding=4), nn.Tanh())def forward(self, x):out1 = self.conv1(x)out = self.res_blocks(out1)out2 = self.conv2(out)out = torch.add(out1, out2)out = self.upsampling(out)out = self.conv3(out)return outclass Discriminator(nn.Module):def __init__(self, input_shape):super(Discriminator, self).__init__()self.input_shape = input_shapein_channels, in_height, in_width = self.input_shapepatch_h, patch_w = int(in_height / 2 ** 4), int(in_width / 2 ** 4)self.output_shape = (1, patch_h, patch_w)def discriminator_block(in_filters, out_filters, first_block=False):layers = []layers.append(nn.Conv2d(in_filters, out_filters, kernel_size=3, stride=1, padding=1))if not first_block:layers.append(nn.BatchNorm2d(out_filters))layers.append(nn.LeakyReLU(0.2, inplace=True))layers.append(nn.Conv2d(out_filters, out_filters, kernel_size=3, stride=2, padding=1))layers.append(nn.BatchNorm2d(out_filters))layers.append(nn.LeakyReLU(0.2, inplace=True))return layerslayers = []in_filters = in_channelsfor i, out_filters in enumerate([64, 128, 256, 512]):layers.extend(discriminator_block(in_filters, out_filters, first_block=(i == 0)))in_filters = out_filterslayers.append(nn.Conv2d(out_filters, 1, kernel_size=3, stride=1, padding=1))self.model = nn.Sequential(*layers)def forward(self, img):return self.model(img)

SRGAN

最后编写训练模块

srgan.py

import argparse
import os
import numpy as np
import math
import itertools
import sysimport torchvision.transforms as transforms
from torchvision.utils import save_image, make_gridfrom torch.utils.data import DataLoader
from torch.autograd import Variablefrom models import *
from datasets import *import torch.nn as nn
import torch.nn.functional as F
import torchos.makedirs("images", exist_ok=True)
os.makedirs("saved_models", exist_ok=True)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=200, help="number of epochs of training")
parser.add_argument("--dataset_name", type=str, default="img_align_celeba_demo", help="name of the dataset")
parser.add_argument("--batch_size", type=int, default=4, 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("--hr_height", type=int, default=256, help="high res. image height")
parser.add_argument("--hr_width", type=int, default=256, help="high res. 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 saving image samples")
parser.add_argument("--checkpoint_interval", type=int, default=50, help="interval between model checkpoints")
opt = parser.parse_args()
print(opt)cuda = torch.cuda.is_available()hr_shape = (opt.hr_height, opt.hr_width)# Initialize generator and discriminator
generator = GeneratorResNet()
discriminator = Discriminator(input_shape=(opt.channels, *hr_shape))
feature_extractor = FeatureExtractor()# Set feature extractor to inference mode
feature_extractor.eval()# Losses
criterion_GAN = torch.nn.MSELoss()
criterion_content = torch.nn.L1Loss()if cuda:generator = generator.cuda()discriminator = discriminator.cuda()feature_extractor = feature_extractor.cuda()criterion_GAN = criterion_GAN.cuda()criterion_content = criterion_content.cuda()if opt.epoch != 0:# Load pretrained modelsgenerator.load_state_dict(torch.load("saved_models/generator_%d.pth"))discriminator.load_state_dict(torch.load("saved_models/discriminator_%d.pth"))# 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))Tensor = torch.cuda.FloatTensor if cuda else torch.Tensordataloader = DataLoader(ImageDataset("../../data/%s" % opt.dataset_name, hr_shape=hr_shape),batch_size=opt.batch_size,shuffle=True,num_workers=3,
)# ----------
#  Training
# ----------for epoch in range(opt.epoch, opt.n_epochs):for i, imgs in enumerate(dataloader):# Configure model inputimgs_lr = Variable(imgs["lr"].type(Tensor))imgs_hr = Variable(imgs["hr"].type(Tensor))# Adversarial ground truthsvalid = Variable(Tensor(np.ones((imgs_lr.size(0), *discriminator.output_shape))), requires_grad=False)fake = Variable(Tensor(np.zeros((imgs_lr.size(0), *discriminator.output_shape))), requires_grad=False)# ------------------#  Train Generators# ------------------optimizer_G.zero_grad()# Generate a high resolution image from low resolution inputgen_hr = generator(imgs_lr)# Adversarial lossloss_GAN = criterion_GAN(discriminator(gen_hr), valid)# Content lossgen_features = feature_extractor(gen_hr)real_features = feature_extractor(imgs_hr)loss_content = criterion_content(gen_features, real_features.detach())# Total lossloss_G = loss_content + 1e-3 * loss_GANloss_G.backward()optimizer_G.step()# ---------------------#  Train Discriminator# ---------------------optimizer_D.zero_grad()# Loss of real and fake imagesloss_real = criterion_GAN(discriminator(imgs_hr), valid)loss_fake = criterion_GAN(discriminator(gen_hr.detach()), fake)# Total lossloss_D = (loss_real + loss_fake) / 2loss_D.backward()optimizer_D.step()# --------------#  Log Progress# --------------batches_done = epoch * len(dataloader) + iif batches_done % opt.sample_interval == 0:# Save image grid with upsampled inputs and SRGAN outputsimgs_lr = nn.functional.interpolate(imgs_lr, scale_factor=4)gen_hr = make_grid(gen_hr, nrow=1, normalize=True)imgs_lr = make_grid(imgs_lr, nrow=1, normalize=True)img_grid = torch.cat((imgs_lr, gen_hr), -1)save_image(img_grid, "images/%d.png" % batches_done, normalize=False)sys.stdout.write("[Epoch %d/%d] [D loss: %f] [G loss: %f]"% (epoch+1, opt.n_epochs, loss_D.item(), loss_G.item()))print(" ")if opt.checkpoint_interval != -1 and (epoch+1) % opt.checkpoint_interval == 0:# Save model checkpointstorch.save(generator.state_dict(), "saved_models/generator_%d.pth" % epoch)torch.save(discriminator.state_dict(), "saved_models/discriminator_%d.pth" % epoch)

原作者构建的srgan.py运行环境在控制台中，这样便于使用但不适合调试。为了方便地调试还是用IDE打开更好一些。

代码中有几个需要注意的问题：

①首先是版本问题，这里要求pytorch为最新版，否则utils的读取可能会出现调包路径错误的问题。

②在dataloader中

dataloader = DataLoader(ImageDataset("../../data/%s" % opt.dataset_name, hr_shape=hr_shape),batch_size=opt.batch_size,shuffle=True,num_workers=3,
)

当出现samle读取数据量为0的错误时，将shuffle=True改为False。
当出现内存错误时，将num_workers=3改为0或直接默认为0。原因是Win系统中不能完美兼容多线程，很多时候因为指定了多个线程内存会莫名的溢出，改为0最保险。

③关于刚开始的默认参数：
batch_size，lr默认数值效果都很好，暂时还未尝试其他取值
epoch因为数据量巨大，应该酌情缩小一些（尝试了只用1k张图片跑epoch50，效果还可以）
checkpoint_interval和sample_interval都是监督训练进度的参数，保存模型和实时对比超分辨率恢复的情况。

实现结果

100张图像，epoch200，sample_interval取500（实际上就是batch跑了多少次，整个除法真上流 ）
保存的部分结果如下：

初始：
500次：

1000次：

2000次：

3000次：