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

REALSR training code #15

Closed ShiinaMitsuki closed 3 years ago

ShiinaMitsuki commented 3 years ago

Is there any plan for the release of the training code on REALSR dataset using IMDN_AS arch?

Thanks.

Zheng222 commented 3 years ago

@ShiinaMitsuki The training processing of IMDN_AS is very similar with IMDN. If you are interested in IMDN_AS, please modify train_IMDN.py for training.

ShiinaMitsuki commented 3 years ago

@ShiinaMitsuki The training processing of IMDN_AS is very similar with IMDN. If you are interested in IMDN_AS, please modify train_IMDN.py for training.

Thanks anyway. I modified one indeed, I trained on REALSR x4 part using the IMDN_x8 model(downsample the input by a factor of 8 to speed up), and I got PSNR ~29, SSIM ~0.82 on the REALSR validation set.

I wonder the performance being acceptable?

Zheng222 commented 3 years ago

@ShiinaMitsuki The downsampling scale factors of fthe RealSR dataset is range from x2 to x4. Thus, I don't recommend x8 downsampling for our model. If you want to further improve the performance, you can train longer time when the learning rate is 2e-4 and 1e-4.