This repository contains a simple PyTorch implementation of 4x SRGAN combined with EDSR, RCAN, MSRN.

Now CPU/OneGPU/DDP versions have been implemented, you can use it freely with bash

1. Linux version 3.10.0-693.el7.x86_64
2. Python 3.8.15
3. Pytorch 1.12.0
4. CUDA Toolkit 11.6
5. cuDNN 8302

Pretrained weight for CNN and GAN also offer Google Drive

1. prepare the dataset DIV2K in whatever directory you want, and assign to args in train.py/pretrain.sh
2. assign the model out path as above too
3. input: bash pretrain.sh
4. if you want to monitor the training, input: bash tensorboard.sh after you assign the model out paht in bash

1. use test.ipynb in MyNet.code, the example has been offered

1. The function torchvision.transforms.Resize used for downsample is not bicubic by default, but has been change to Bicubic now
2. Training and Validating Dateset is DIV2K
3. Data augmentation: random crop -> random rotation in 90 * k degree/flipping, you can change it in MyNet.code.dataset.py

1. You can assgin your own lr_scheduler in net.solver.init
2. Train on RGB, but validate on Y. However, validate on RGB is also offer in MyNet.code.net.solver.test
3. D_train require d_real_loss and d_fake_loss more than 0.4 to aviod D too strong, you can change the value in train.py's D_threshold !!!
4. Hyper parameter K in GAN is not supported(default==1) because it is useless in practice, you have to recode net.solver.run/run_resume if you want
5. pretrain G with 1e-4 lr for 2000 times, and then 1e-5 lr for 18000 times
6. run G and D with 1e-5 lr for 10000 times
7. the more iteration you train in pretrain and run, the better G is, if you want know more, read about SRGAN or ESRGAN

1. PSNR and SSIM calculations are only on the y channel
2. Use float32 for PSNR calculation and uint8 for SSIM calculation

Please feel free to contribute to the project, I am pleasured to answer any questions

This code is built on SRGAN (PyTorch). Thank a lot to the authors for sharing codes!