WIP: Added support for timm in unet

[madhavajay]

This PR attempts to add timm models to the unet_learner, as per conversations during Live Coding 17: https://forums.fast.ai/t/live-coding-17/97166

I have several issues so far:

1. I don't know if theres a way to get the cut preferences from timm, I have added some code which tries to match them to fastai model types as a backup but they don't seem to match up so perhaps this is pointless and instead we either get them somewhere else or rely on manual user input

2. I tried training with them, but it seems like they use up a tonne of memory meaning my batch size can only be 1, not sure whats going on but something seems wrong especially considering i tried a smaller timm model resnet18 than my default unet fastai model resnet32.

I am sure I have done something wrong, and would appreciate some direction on what to do next.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[madhavajay]

So, I tried training it just to make sure its still working.

I got this far before my paperspace machine shutdown:

So I guess its definitely training but extremely slowly. For comparison I can do the same dataset, resnet34 with batch size 4 and get epochs of about 6 minutes on that free paperspace gpu. So assuming the architecture was equally complex and the batch size was 4 times lower that should only take 24 minutes per epoch.

It seems like when the model is allocated theres only 7% gpu memory in use, and then once training with 1 batch starts it goes to 87+%

I guess perhaps I don't understand the model cutting code and how to use it with timm models. Any advice on how to debug this?

[madhavajay]

Okay, I have changed the code use timm.create_model features_only=True.
So far it seems to be training a convnext_tiny