Automatically download / upload models with hugging face, torch hub

[Nic-Ma]

Is your feature request related to a problem? Please describe.
As MONAI bundle features are ready to use, it would be nice to manage bundle models with hugging face and torch hub, etc.
Like: https://huggingface.co/MONAI/example_spleen_segmentation

So it's necessary to develop features to automatically download / apply models from the hubs, may support bundle package, checkpoint, torchscript model. And for future steps, we can think about how to review new models and execute CICD before approving or uploading to the hub.

[Nic-Ma]

@yiheng-wang-nv is investigating requirements, ideas, solutions, will prepare a proposal soon.

Thanks.

[Nic-Ma]

Hi @yiheng-wang-nv ,

In my simple mind, I think maybe we can start from the download features and think about below initial ideas first:

1. Use cases to download with CLI python -m monai.bundle download XXX:
   1.1. download the whole bundle package
   1.2. download the model weights
   1.3. download the TorchScript model
   1.4. download any file, refer to: https://huggingface.co/docs/hub/how-to-downstream
2. Data source:
   2.1. bundle packages on hugging face
   2.2. (P1) bundle packages on NVIDIA NGC
   2.3. (P1) bundle packages on amazon AWS
   (I think torch hub is just APIs, no storage)
3. Support 3rd APIs for advanced features:
   3.1. use torch.hub.XXX, refer to: https://github.com/huggingface/transformers/blob/main/hubconf.py
   3.2. (P1) use hugging face API, refer to: https://huggingface.co/docs/hub/adding-a-library
4. Update the tutorial example to show features.

Thanks.