search

asteroid-team / asteroid

The PyTorch-based audio source separation toolkit for researchers
https://asteroid-team.github.io/
MIT License
2.27k stars 423 forks source link

Problems on building model : DCCRN on DNS-Challenge dataset #446

Closed realkris closed 3 years ago

realkris commented 3 years ago

This project is really helpfull for noobs like me ! I want to make a DCCRN model(on DNS-Challenge official dataset) for class project as the noise cancellation module. Here's What I've done now. I modified the model.py script in egs/dns-challenge/baseline according to #278, but I have several questions that confused

  1. In function distance(), where does the "F" come from?
  2. how can I change the channels of DCCRN? I saw there's a "_dccrn_architectures.py " seems like a config of DCRNN under asteroid/masknn, but I didn't find where it is called in model architecture.
  3. Do we only have DCCRN-CL as the architecture to use? is that because DCCRN-CL is better than the "DCCRN - E/R/C"?

So basically I just modified the model.py scripy and dataset folder in baseline/conf.yml. Will that be enough? My English is not good, maybe I missed something? or could you guys just give me an example.(I will train on the official DNS dataset) Thank you for your help!

jonashaag commented 3 years ago

I think it’s easiest to start from this recipe: https://github.com/asteroid-team/asteroid/tree/master/egs/librimix/DCCRNet

F is an alias convention for torch.nn.functional.

CL is the only variant available because nobody implemented the others yet.

What modification to the channels do you want to make? Input, output, or inner (ie. model size)?

realkris commented 3 years ago

Thank you for your reply! yes, model size, I want to reduce the size of model cause I want to run it on portable devices. is it correct to change the kernel_num from like [32, 64, 128, 256, 256, 256] to [16, 32, 64, 128, 256, 256]?

jonashaag commented 3 years ago

You’ll have to try. Nobody knows. But doesn’t look crazy wrong.

realkris commented 3 years ago

OK, thank you for your help, I'll continue trying it!