SRNext

This respository contains code to train & eval SRNext - A (potential) state of the art in lightweight super-resolution.

Original	SwinIR	SRNext

Similar results, but up to 1.6x faster with 60% less utilised memory!

Compared with SOTA

PNSR

Model	Set14	BSD100	URBAN100	MANGA109
IMDN	27.8040	24.0524	27.2880	23.4331
CARN	28.0326	24.0659	27.3436	23.4023
SwinIR	28.2085	24.0346	27.7790	23.4692
SRNext	28.0364	24.1766	27.1437	23.7909

SSIM

Model	Set14	BSD100	URBAN100	MANGA109
IMDN	0.7508	0.7002	0.8201	0.8176
CARN	0.7490	0.6730	0.7945	0.7885
SwinIR	0.7685	0.7080	0.8397	0.8297
SRNext	0.7677	0.7103	0.8263	0.8338

Report

Find out more details in the report

Quick Start

Install libraries

pip install -r requirements.txt

Start training (will go on forever, must Ctrl+C).

python3 train.py

Evaluate SRNext

python3 eval.py

Inference SRNext on an image

python3 inference.py sample.png sample_out.png 

Setup

Datasets are primarily taken from kaggle: FLICKR2k DIV2k and combined manually.

Simiarly, test sets are from here.

Structure

'Models' and 'Dataset' definitions are stored in the 'archs' directory, eg: The SRNext architecture is defined at archs/models/srnext.py

'Methods' takes models and trains them. They are very generalised don't initialise models, eg: methods/bootstrap.py

'Experiments' combines a method and their models. Eg, experiments/bootstrap_unext.py will take an untrained unext model and pretrained realesr model, then train unext to match the outputs of realesr.

Finally, to actually start the experiment, it is imported in train.py than ran with python.