search

Zheng222 / IMDN

Lightweight Image Super-Resolution with Information Multi-distillation Network (ACM MM 2019, Winner Award of ICCVW AIM 2019 Constrained SR Track1&Track2)
418 stars 69 forks source link

About the performance in x2 model #20

Open NJU-Jet opened 3 years ago

NJU-Jet commented 3 years ago

Thanks for your excellent work. But I hava a doublt about the results on x2 model. I can't get the results as you report in the article, so should I use L2 loss after 1000 epochs? Thanks!

Zheng222 commented 3 years ago

@NJU-Jet Hello, it doesn't need to use L2 loss. I recommend that training more iterations on larger learning rates, like 2e-4 and 1e-4. The performance improvement is mainly in these two learning rates.