search

lucidrains / audiolm-pytorch

Implementation of AudioLM, a SOTA Language Modeling Approach to Audio Generation out of Google Research, in Pytorch
MIT License
2.33k stars 249 forks source link

Eos handling #200

Closed LWprogramming closed 1 year ago

LWprogramming commented 1 year ago

Prevent eos from being generated after going through all quantizers in a timestep, instead of preventing eos except for the last quantizer within a timestep. For issue #199

lucidrains commented 1 year ago

oh dang! I think you are right 🙏

have you tested this?

lucidrains commented 1 year ago

ok merging, 'lgtm', as they say

LWprogramming commented 1 year ago

oh dang! I think you are right 🙏

have you tested this?

Ah yep! I tried it on the same original (non-trimmed) input data configuration and it failed in the same way without this and I was able to generate some sort of output with it.