search

sanghyun-son / EDSR-PyTorch

PyTorch version of the paper 'Enhanced Deep Residual Networks for Single Image Super-Resolution' (CVPRW 2017)
MIT License
2.42k stars 668 forks source link

Reproduces lower PSNR than reported on the paper #306

Open DongNaeSwellfish opened 3 years ago

DongNaeSwellfish commented 3 years ago

I trained the EDSR baseline x 2 and x4 on DIV2K and tested no URBAN100, and the result is poor. For x 2, I got 31.949, and for x4, I got 25.62, but the paper reported it to be 32.93 and 26.64 each. Is there anyone got lower result like me and solved this problem?

REN-HT commented 3 years ago

I trained the EDSR baseline x 2 and x4 on DIV2K and tested no URBAN100, and the result is poor. For x 2, I got 31.949, and for x4, I got 25.62, but the paper reported it to be 32.93 and 26.64 each. Is there anyone got lower result like me and solved this problem?

so, do you have solved this problem?

DongNaeSwellfish commented 3 years ago

Yes. I missed the detailed settings. The EDSR model that produces the reported score has larger parameters compared to other models. The channel size and num_resblocks were 256 and 32 each, while other models usually use 64 and 16 each.

2021년 7월 15일 (목) 오후 5:42, Brand Ren @.***>님이 작성:

I trained the EDSR baseline x 2 and x4 on DIV2K and tested no URBAN100, and the result is poor. For x 2, I got 31.949, and for x4, I got 25.62, but the paper reported it to be 32.93 and 26.64 each. Is there anyone got lower result like me and solved this problem?

so, do you have solved this problem?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/sanghyun-son/EDSR-PyTorch/issues/306#issuecomment-880512644, or unsubscribe https://github.com/notifications/unsubscribe-auth/AGH7ZMHDSHKKL5KUUAAEC43TX2NO5ANCNFSM44APNOUQ .