GitHub - luissen/ESRT (original) (raw)

Efficient Transformer for Single Image Super-Resolution

Update

#######22.03.17########

The result images of our method are collected in fold "/result".

Environment

• pytorch >=1.0
• python 3.6
• numpy

Model

The overall architecture of the proposed Efficient SR Transformer (ESRT).

Efficient Transformer and Efficient Multi-Head Attention.

Train

• dataset: DIV2K
• prepare
  Like IMDN, convert png files in DIV2K to npy files:
  python scripts/png2npy.py --pathFrom /path/to/DIV2K/ --pathTo /path/to/DIV2K_decoded/
• Training

python train.py --scale 2 --patch_size 96 python train.py --scale 3 --patch_size 144 python train.py --scale 4 --patch_size 192

If you want a better result, use 128/192/256 patch_size for each scale.

Test

Example:

• test B100 X4

python test.py --is_y --test_hr_folder dataset/benchmark/B100/HR/ --test_lr_folder dataset/benchmark/B100/LR_bicubic/X4/ --output_folder results/B100/x4 --checkpoint experiment/checkpoint/x4/epoch_990.pth --upscale_factor 4

Visual comparison

[](/luissen/ESRT/blob/main/figs/visual images-v2.png)
The visual comparison.