search

njulj / RFDN

Residual Feature Distillation Network for Lightweight Image Super-Resolution
MIT License
348 stars 43 forks source link

RFDB is difference with figure in the paper. #3

Open GuideWsp opened 3 years ago

GuideWsp commented 3 years ago

Finally, I found the code of the RFDB is not the same as the paper.

njulj commented 3 years ago

The only difference is that we did not use big residual connections in RFDB, which we found is slightly better in the challenge

GuideWsp commented 3 years ago

Yes, I had try it as the paper, finally I got a 30.47dB on DIV2K-val(YCbCr), the psnr of your model is 30.49dB, they are very similar. I just train it by L1, not finetune by L2. Whatever, it is a very good job.

scutlrr commented 3 years ago

Yes, I had try it as the paper, finally I got a 30.47dB on DIV2K-val(YCbCr), the psnr of your model is 30.49dB, they are very similar. I just train it by L1, not finetune by L2. Whatever, it is a very good job.

I also used the provided RFDN_AIM.pth file for testing, and the test result on DIV2K_val (only Y channel) was 30.517db. Maybe it is caused by the different calculation methods of psnr.

aiyunhug commented 2 years ago

唯一的区别是，我们在RFDB中没有使用大的残差连接，我们发现这在挑战中略好一些。

why is 'self.remaining_channels = in_channels' in class RFDB？

Liiiiaictx commented 1 month ago

@scutlrr 大佬，这个RFDN怎么基于EDSR框架训练的，需要把这个仓库的文件拷贝到EDSR仓库吗？其次RFDN的环境配置是什么样的？

Liiiiaictx commented 1 month ago

@aiyunhug 大佬，这个RFDN怎么基于EDSR框架训练的，需要把这个仓库的文件拷贝到EDSR仓库吗？其次RFDN的环境配置是什么样的？