Unstable training of DPRNN-TasNet

tky823 / DNN-based_source_separation

A PyTorch implementation of DNN-based source separation.

286 stars 50 forks source link

Closed tky823 closed 3 years ago

tky823 commented 3 years ago

The network wouldn't be trained when

Layer normalization is applied before the bottleneck convolution.
Batch size / GPU is small (e.g. 1).
- This phenomenon is observed in Conv-TasNet.

No bottleneck convolution seems to be better.

tky823 commented 3 years ago

Related to #23

tky823 commented 3 years ago

The bias of bottleneck convolution?

tky823 commented 3 years ago

Possible causes ...