Image super-resolution using IMDN (information multi-distillation network) modified with multi-head structure with pytorch
DATASET : DIV2K, Flickr2K
You should download dataset, then run png2npy.py to convert png to npy
python png2npy.py
- DIV2K
- DIV2K_train_HR
- DIV2K_train_LR_bicubic
- DIV2K_valid_HR
- DIV2K_valid_LR_bicubic
- Flickr2K
- Flickr2K_HR
- Flickr2K_LR_bicubic
- Flickr2K_LR_unknown
- DIV2K_decoded
- DIV2K_HR_train
- DIV2K_HR_valid
- DIV2K_LR_bicubic_train
- X3
- DIV2K_LR_bicubic_valid
- Flickr2K_decoded
- Flickr2K_HR
- Flickr2K_LR_bicubic
- X3
then run train.py
python train.py
without pretrained model
python train.py --resume checkpoint/model.pth
with pretrained model
convert model to demo.py readable format
python load_model.py
then run demo.py
python demo.py
to run video demo
python video_SR.py
LR | SR | HR |
---|---|---|
![]() |
![]() |
![]() |