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

Separate transformer and trainer checkpoint load logic #214

Closed LWprogramming closed 12 months ago

LWprogramming commented 12 months ago

Since we no longer want to instantiate multiple Trainer instances.

Also fixes a device-related bug in the process (see last commit):

This was covered up by previous usage where we'd instantiate trainers for each transformer (semantic/coarse/fine). Now that we have to be able to load transformers without their corresponding trainers, because we can only load one accelerator at a time, the code in trainer that checks wrapper device no longer applies and we get device mismatches. This commit fixes that

--

I tested this code and was able to get a small training run of a few hundred steps working so it works e2e!

lucidrains commented 12 months ago

lgtm!