Codecov Report

Merging #5535 (5ec639f) into master (eb6f93a) will decrease coverage by 0.25%.
Report is 14 commits behind head on master.
The diff coverage is 75.00%.

@@            Coverage Diff             @@
##           master    #5535      +/-   ##
==========================================
- Coverage   75.43%   75.19%   -0.25%     
==========================================
  Files         711      711              
  Lines       65757    65757              
==========================================
- Hits        49607    49449     -158     
- Misses      16150    16308     +158     

Flags with carried forward coverage won't be shown. Click here to find out more.

... and 4 files with indirect coverage changes

📣 Codecov offers a browser extension for seamless coverage viewing on GitHub. Try it in Chrome or Firefox today!

Thanks, but we still have an issue...

This is weird. I cannot reproduce the error with the same PyTorch version...

Is this operation actually used?

Are you talking about the DefaultFrontend with WPE and beamforming? I think we usually use the S2T function (SE+ASR) instead of ASR only, where the new NeuralBeamformer (espnet2/enh/separator/neural_beamformer.py) is used.

We still use the DefaultFrontend for FBank feature extraction, but usually not with beamforming.

I mean the matrix inverse function, not wpe or beamformer.
I remember that there are several implementations for the complex matrix inverse operation.

Because it is using the pytorch_wpe package for WPE (where the CI error occurs), the matrix inverse operation is conducted based on torch_complex. And yes, the error message shows the torch_complex package is used for matrix inverse. Both packages are not maintained in our codebase.

Got it.
Is there a way to avoid the issue of going through torch_complex by copying them inside espnet and using the other matrix inverse in the future?

I want to remove such dependencies in the future.

I think so. It is what I did before, and probably we can just replace

from espnet.nets.pytorch_backend.frontends.dnn_beamformer import DNN_Beamformer
from espnet.nets.pytorch_backend.frontends.dnn_wpe import DNN_WPE

with the new implementations in espnet2

from espnet2.enh.layers.dnn_beamformer import DNN_Beamformer
from espnet2.enh.layers.dnn_wpe import DNN_WPE

Sounds good!
With this, can we remove these dependencies?
If so, we may consider it in the other PRs. This will simplify things.

I think at least we can remove pytorch_wpe. It was only used in espnet/nets/pytorch_backend/frontends/dnn_wpe.py. I have verified locally that replacing espnet1 implementations with espnet2 ones can pass the tests.

Nice!
This is a good part of having a unit test!
We can safely replace it :)

torch_complex is still now used in Enh since some complex-valued functions have not been implemented yet in torch.

OK, it passes the test!
Can you also remove pytorch_wpe, then?

Sure.

Can you remove it at https://github.com/espnet/espnet/blob/master/setup.py#L42?

Yes, I already removed it in the latest commit.

Thanks a lot!
This was a difficult bug, but it brought an outcome to remove the old dependency!
Very good!