search

JingyunLiang / RVRT

Recurrent Video Restoration Transformer with Guided Deformable Attention (NeurlPS2022, official repository)
https://arxiv.org/abs/2206.02146
Other
354 stars 33 forks source link

Reproducing the Video Denoiser #25

Open gauenk opened 1 year ago

gauenk commented 1 year ago

Hello RVRT Team :wave: Thank you again for your great code.

I have followed your paper's instructions to train my own RVRT model. However, my network's quality does not match the network you share online, nor the numbers you report.

Maybe I have missed something on my end, but I want to know if you have a training script so I can inspect all the details myself?

Thank you.

gauenk commented 1 year ago

I am gently following up on the previous comment.

Dawson0813 commented 1 year ago

I caught the same problem with you, even though the iterator number more than 20000 times, the loss almost stay the same(8e-2). Thus the infference results of the new-trained model is still far from the pre-trained one.

gauenk commented 1 year ago

As a researcher with limited GPUs, your comment means more to me than you know.

Dawson0813 commented 1 year ago

As a researcher with limited GPUs, your comment means more to me than you know.

Thank you for your kindly reply too. I wonder whether you have solved above issues. Last week I used a small trainning-sets to verify the RVRT network, but both the training loss and the test PSNR still stay the same.

Dawson0813 commented 1 year ago

The test results of author's pre-trained model is realy good, thanks again for their great work. I am sure that I have missed some key procedures in my training too, and I am also eagerly looking forward to get some more details from the RVRT team.

gauenk commented 12 months ago

My issue is still unresolved.