novwaul / SwinCIR

Proposed model on Individual Study (Lab. Intern)

SwinCIR: SwinIR with multi-resolution connections

4x Result

PSNR

Dataset	Bicubic	SwinCIR
Set5	28.648	32.460
Set14	26.230	28.793
Urban100	23.220	26.329

Set14

Bicubic	SwinCIR	GT

Train Setting

Item	Setting
Train Data	DIV2K, Flickr2K
Preprocess	[-1,1] Normalization
Random Transforms	Crop {DIV2K(64x64), Flickr2K(64x64)}, Rotation {90 degree}
Validation Data	DIV2K
Test Data	Set5, Set14, Urban100
Scale	4x
Optimizer	Adam
Learning Rate	2e-4
Scheduler	Cosine Annealing with last LR of 1e-6
Trained Iterations	5e5
Loss	L1
Batch	4 {2 for each GPU, total 2 GPUs are utilized}