search

ckkelvinchan / RealBasicVSR

Official repository of "Investigating Tradeoffs in Real-World Video Super-Resolution"
Apache License 2.0
906 stars 136 forks source link

train code #11

Closed sunyclj closed 2 years ago

sunyclj commented 2 years ago

Hello,when will the train code be released？I hope to get the x2 pre-trained model through train.thanks.

ckkelvinchan commented 2 years ago

Hi @sunyclj, the training code is currently under review in MMEditing. It will be available by the end of this month. Please stay tuned!

sunyclj commented 2 years ago

@ckkelvinchan ok,thank u.

Ha0Tang commented 2 years ago

@ckkelvinchan Hi how about the code?

ckkelvinchan commented 2 years ago

Hello, the training code is actually on MMEditing now. But since the latest version of MMEditing is not released, we cannot use mim install mmedit and train directly.

You can refer to MMEditing for the training for now. After the release of the latest version, I will update this repo, thanks for your patience and sorry for the delay.

ckkelvinchan commented 2 years ago

Hello @Ha0Tang @sunyclj, the training code has been added. Please try and let me know if there are any problems.

sunyclj commented 2 years ago

@ckkelvinchan thanks.when running 'mim train mmedit configs/realbasicvsr_wogan_c64b20_2x30x8_lr1e-4_300k_reds.py --gpus 3 --launcher pytorch', error occurred, AssertionError: Please install av to use video compression.what is av,please.

ckkelvinchan commented 2 years ago

Please use pip install av for installation. I will add it to README. Thanks for your reminder.

sunyclj commented 2 years ago

@ckkelvinchan I wanted to train an x2 model, would it work if I just changed scale = 2 in the realbasicvsr_wogan_c64b20_2x30x8_lr1e-4_300k_reds.py file?

ckkelvinchan commented 2 years ago

No, but you can modify BasicVSR so that it works for 2x upsampling. More specifically, you can remove one upsampling layer here for 2x upsampling.

If you would like to modify the code, you will need to clone MMEditing and follow the installation instruction here.

sunyclj commented 2 years ago

@ckkelvinchan Yes.I modified some code and parameters,got x2 model.I have a question，in folder basicvsr_net.py ，Why do you upsample LR image to get base and out + = base? thanks. 1642123148409_F5D41874-1EC4-49c7-A070-6CE44576E63A

ckkelvinchan commented 2 years ago

This is because we want to learn only the residue, since the input LR image should provide enough low-frequency information, and hence we do not need to learn it from scratch.

zhixiongzuo commented 2 years ago

Hi @sunyclj , how can you get x2 model? I comment upsample2 in basicvsr_net.py and change scale_factor to 2, but it didnot work.