search

desh2608 / gss

A simple package for Guided source separation (GSS)
MIT License
105 stars 13 forks source link

do not chunk when performing beamforming #45

Open popcornell opened 4 months ago

popcornell commented 4 months ago

as pointed out by @kamo-naoyuki, there are currently problems when doing the beamforming due to the fact that the a-posteriori SNR reference channel selection is performed for each freq chunk independently (and this requires a reduction over the freq axis). The easy fix is to avoid to chunk frequencies when doing the beamfoming. Hopefully it is still ok and will not lead to OOMs

popcornell commented 4 months ago

@kamo-naoyuki, did you had a chance to try ? For me the results are the same as with the old code. But may it be because my GPUs are 40GB VRAM.

@desh2608 can you merge ? I confirmed this does not impact the results.

desh2608 commented 4 months ago

I am concerned that this may cause GPU OOMs when the segment is long and/or we are using a large number of channels. It may be better to have a CLI option to set bf chunking on/off (turned on by default). Could you make this change?

popcornell commented 4 months ago

Yes but then we need to also disable the channel selection inside the beamformer when the option is disabled.

Or IDK beamforming part needs to be changed quite a lot to accomodate for chunking but only before the channel.selection