Reshaping issue about speech-separation HOT 9 OPEN

Hi,

Thanks for pointing out the mistake. It was fixed when the repo was merged to asteroid here, somehow it was not fixed here. But I did not see any significant improvement in terms of overfitting unfortunately.

from speech-separation.

Thanks for the quick response @vitrioil ! It seems another similar bug is repeated here and also in the asteroid. Here, we are reshaping (N,298,2x257xself.num_person) into (N,2,298,257,self.num_person). Instead of view(N, 2, 298, 257, self.num_person), it will be more appropriate to do view(N,298,2,257,self.num_person).transpose(1,2).

from speech-separation.

I see... we are mixing the axis at the last layer. Thank you again for pointing this out! I will fix this.

from speech-separation.

I have raised a bug in asteroid as well. Another thing, this line needs to be commented out, I guess.

from speech-separation.

A possible fix:
In line #187 of models.py, replace output_layer = output_layer.view(batch_size,-1,height,1) with output_layer = output_layer.permute(0, 2, 1, 3).contiguous().view(batch_size, height, -1, 1).permute(0, 2, 1, 3)

from speech-separation.

please let know if you manage to overcome the overfitting problem with this fix. thanks!

from speech-separation.

Here input channel expected by self.conv1 is 512 so transposing the channels so that we get the correct expected channel, otherwise channels will still be 1. I don't have access to any GPU now but if I do train it and see an improvement I will let you know.

from speech-separation.

Hi,
Firsly, thanks for the code.
Is there any improvement in the overfitting problem?
@vitrioil

from speech-separation.

Hi @MordehayM,
I still haven't tried re-training, so currently the result is the same

from speech-separation.