Add integration with HF Hub

[qubvel]

Description

Hi! Amazing repo!

I made a draft with PytorchModelHubMixin for a few models in the repo. This mixin adds a few methods to make it possible to save/load models and share them with the HF hub.

Here is a short example:

from monai.networks.nets import AHNet

net = AHNet()

# save to a local folder
net.save_pretrained("qubvel-hf/ahnet")

# push to HF hub
net.push_to_hub("qubvel-hf/ahnet")

# load form local folder or HF hub
net = AHNet.from_pretrained("qubvel-hf/ahnet")

I believe the repo and community may benefit from sharing pretrained models built with this framework! Let me know if it aligns with your roadmap, I will be happy to help you with implementation or to complete the PR!

Types of changes

• [x] Non-breaking change (fix or new feature that would not break existing functionality).
• [ ] Breaking change (fix or new feature that would cause existing functionality to change).
• [ ] New tests added to cover the changes.
• [ ] Integration tests passed locally by running ./runtests.sh -f -u --net --coverage.
• [x] Quick tests passed locally by running ./runtests.sh --quick --unittests --disttests.
• [ ] In-line docstrings updated.
• [ ] Documentation updated, tested make html command in the docs/ folder.

[qubvel]

Hey! I would appreciate any comments on this 🤗 ping @ericspod @KumoLiu for visibility

[ericspod]

Hi @qubvel Thanks again. Please double check places like this in setup.cfg have the updated information as your code mentions needed safetensors. Is it possible to perhaps subclass these networks as well? I am hesitant to modify existing definitions as a general rule even though things like the model weights are unchanged. For example instead of changing AHNet we would have class AHNetHub(AHNet, MonaiHubMixin): pass. I'd like some feedback from others on this as well.

[qubvel]

@ericspod thanks for the review, I will address the comments! At the moment it's a raw draft to know how you feel about this integration in general. Any comments on how to make it smooth are appreciated!

I'm not sure I got the point here, can you provide more details?

Is it possible to perhaps subclass these networks as well? I am hesitant to modify existing definitions as a general rule even though things like the model weights are unchanged.