lucidrains / audiolm-pytorch

Implementation of AudioLM, a SOTA Language Modeling Approach to Audio Generation out of Google Research, in Pytorch
MIT License
2.36k stars 255 forks source link

Use correct eos id when masking out #122

Closed LWprogramming closed 1 year ago

LWprogramming commented 1 year ago

By the way, why is keep_eos hardcoded as False in some places here? I searched through the blame and it used to use include_eos_in_output but it seems like there was something about maybe the hierarchical part here

lucidrains commented 1 year ago

@LWprogramming thank you! 🙏

so i used to keep the eos, and then taught the subsequent transformers to ignore the eos token. but it got to be too complicated, as i had to make sure the eos id is consistent across all wrappers, and do a bunch of self attention masking. so i redesigned it to simply remove the eos altogether