真实数据超分三大方法（BSRGAN、Real ESRGAN、SwinIR）训练细节整理

本文整理单帧真实数据超分目前最有效的三个方法BSRGAN、Real ESRGAN、SwinIR（ 方法详解）的训练细节，方便训练时查找。SwinIR这里只记录了其在Real world SR上的相关配置。

1. 训练代码

• BSRGAN和SwinIR

  https://github.com/cszn/KAIR

• Real ESRGAN

  https://github.com/xinntao/Real-ESRGAN/blob/master/docs/Training.md

2. 训练数据

	数据（只需要HR图像）	退化模型
BSRGAN	DF2K (DIV2K and Flickr2K) , WED and 2,000 face images from FFHQ	BSRGAN
Real ESRGAN	DF2K (DIV2K and Flickr2K) + OST	Real ESRGAN
SwinIR	DF2K (DIV2K and Flickr2K) + OST	BSRGAN

1. DIV2K(800 training images):http://data.vision.ee.ethz.ch/cvl/DIV2K/DIV2K_train_HR.zip
2. Flickr2K(2650 images): https://cv.snu.ac.kr/research/EDSR/Flickr2K.tar
3. OST(10324 images for sky,water,grass,mountain,building,plant,animal): https://openmmlab.oss-cn-hangzhou.aliyuncs.com/datasets/OST_dataset.zip
4. FFHQ（Flickr-Faces-HQ Dataset）：https://github.com/NVlabs/ffhq-dataset
5. WED(Waterloo Exploration Database: New Challengesfor Image Quality Assessment Models, 4,744 images)：https://ece.uwaterloo.ca/~k29ma/exploration/

3. 训练过程

1. 先用 L1 loss训练一个PSNR导向的模型；
2. 以第一阶段的模型参数初始化生成器参数，训练一个Gan;

4. 训练超参数

• PSNR training stage

  	lr	iterations	lr is halved	loss
  BSRGAN	1e-4	2000k	[200K, 400K, 600K, 800K, 1000K]	L1
  Real ESRGAN	2e-4	1000K	无	L1
  SwinIR	2e-4	1000K	[500K,800K,900K,950K]	L1

• GAN training stage

  	lr	iterations	lr is halved	loss
  BSRGAN	G:5e-5; D: 5e-5	1600K	[800k]	1* L1+1* perceptual loss + 0.1*gan loss
  Real ESRGAN	1e-4	400K	无	1* L1+1* perceptual loss + 0.1* gan loss
  SwinIR	G: 1e-4; D:1e-4	600K	[400K,500K,550K,575K]	1* L1+1* perceptual loss + 0.1* gan loss