Native support to load & freeze embeddings for Tabular Learner

I absolutely enjoy the fast.ai’s Tabular Learner API besides Vision & Text with native support to Transfer Learning. Thanks a lot to the team for building and making it open source.

Is there a plan to natively support to load embeddings from a file / model, freeze / unfreeze them while training in v2? I think this would be quite helpful as it is an effective way to transfer learn on tabular datasets.

I saw a couple of approaches by fastai-users to achieve this, but what would be an optimal way to achieve the same with current APIs?

Could you show an example notebook of what you’re looking to achieve - or a link to someone else that’s done it?

@jeremy see the discussion here Tabular Transfer Learning and/or retraining with fastai

Many thanks - I’ll bookmark that for when we get to this.

3 Likes