refactor: load all pretrained models from HuggingFace#6
Conversation
joanise
left a comment
joanise
left a comment
There was a problem hiding this comment.
OK, posting my observations during my early review, and moving to something else for now.
I had a pretty big issue trying to run everyvoice train text-to-wav config/everyvoice-text-to-wav.yaml --mode first in a GPSC-C job: the download stalled a number of times, I had to kill it a few times and relaunch before it actually worked. Can we provide a pre-download mechanism that maybe could be run on a machine without GPU?
| description="HuggingFace repo ID for the PLBERT text encoder.", | ||
| ) | ||
| checkpoint_filename: str = Field( | ||
| default="step_1000000.t7", |
There was a problem hiding this comment.
The file in the repo is actually called 1100000, not 1000000.
| return str(local_path) | ||
| from huggingface_hub import hf_hub_download | ||
|
|
||
| return hf_hub_download(repo_id, filename=filename) |
There was a problem hiding this comment.
With the typo in the file name, this gave me a 404 with a long stack trace. Can we do something friendlier? Not immediately required, but would be nice.
2 participants
Instead of loading locally defined pretrained models, we load them from huggingface. If accepted, we should also scrub the *.t7 and *.pth files from this repo too, to save bloat in the git history.