Fix Chefer and add standardized API

[Logiquo]

This PR furthur enhance the interpretability API

1. Move interpretability related API to a separate ABC Interpretable for model want to be interpretable to inherite, this make the BaseModel ABC cleaner
2. Create a new API called CheferInterpretable for Chefer specific API
3. Remove Chefer model specific attribution implementation, and use Chefere specific API to fetch information instead.

Model

• Transformer
• StageAttn
• ViT (will be defered to later PR)

Methods

• Chefer

[jhnwu3]

everything else lgtm. I like the idea of defining Interpretable abstractions as technically it is different.

I also think there are technically 3-4 types of "Interpretable" abstractions as I'm categorizing them in the literature:

1. Black-box interpreters, which if we properly implement them, technically shouldn't need to have access to embeddings at all. We may need to revisit SHAP and LIME's design later.
2. Counterfactual based, GIM, DeepLift, and Integrated Gradients all require hooks + embedding access. I guess this is our Interpretable abstraction atm
3. Attention-based (CheferInterpretable)
4. Convolution-based (GradCAM, CAM, etc.) - not yet implemented, but are very much related

[jhnwu3]

Just a clarification question, I guess it's because our TorchVision wrapper doesn't/can't support a CheferInterpretable here due to technically not being a static model?

https://github.com/sunlabuiuc/PyHealth/blob/master/pyhealth/models/torchvision_model.py

[Logiquo]

Due to i haven't had time to modify the ViT yet🤣, technically nothing stop it from happening. but given that we are working on interpretability project more, i want to fix the most relavent model first.

[jhnwu3]

lgtm, we can iterate further on the API later when we finalize some of these things.