Freezing layers from CLI

[Lihi-Gur-Arie]

Thank you for this amazing project.
Does anybody know if there is a way to freeze layers for transfer learning, directly from the CLI (when running the train.py)?
Thank you!

[rwightman]

@Lihi-Gur-Arie there is nothing implemented, the API would be somewhat complicated for the fully granular case (ie at individual layer granularity, would have to pass a loooong list of names via command line or a text file).

Using the timm grouping interface I could add an index based freeze, but one would need to be familiar with the grouping for each network to know what the index numbers mean, I implemented something like this was implemented in open_clip for timm backbones

https://github.com/mlfoundations/open_clip/blob/37b729bc69068daa7e860fb7dbcf1ef1d03a4185/src/open_clip/timm_model.py#L85-L115