This wiki entry details some known issues, FAQ and what features are planned to be worked on neosr.

FAQ

Does neosr support other GPUs?

No. Currently only NVIDIA gpus are supported. There's two reasons for it:

• ROCm support is very limited at the moment, only Linux works somewhat and on few gpus.
• I don't have one of the supported gpus to test if neosr is running correctly.

Does neosr support DDP (distributed training)?

Maybe. I currently don't have multiple GPUs for testing the DDP code, so it may or may not work. If you'd like to help testing, please send error traces in issues or on discord, it will be very helpful.

I'm getting memory issues on windows!

If you are on Windows, increase your page file size: View advanced system settings > Advanced > Performance Settings > Advanced > Change > Uncheck "Automatically manage paging file size for all users". Then set "Custom Size" and choose a high number (>10gb). If you still get issues, try setting num_worker_per_gpu manually.

I'm getting NaN during training!

This could be due to multiple reasons. In general, use the default configuration templates and it should work. If it is not, try changing the discriminator and learning rates. A 'safe' setting is to use unet with learning rate of 1e-4. If that doesn't work, try disabling AMP. If you're really stuck, please report it on our discord.

Known Issues

• Learning rate changes after starting training are not working as it should. Use Schedule-Free to avoid changing learning rates, or ConsideAnnealing instead of step scheduler.

Roadmap

Bellow is a TODO list, in order of importance and the current status. Feel free to send PRs in case you feel like contributing!

• [todo] Switch dataloader to NVIDIA DALI.

• [todo] Add VQGAN degradation learning from VQD-SR.

Low priority:

• [todo] Add possibility to call multiple dataroot_* datasets.

• [todo] The option total_iter is reading based on Epochs instead of Iters.