Repository navigation
Automatically download / upload models with hugging face, torch hub #4042
Description
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.
Metadata
Metadata
Assignees
Type
Projects
- Status
Activity
Nic-Ma commentedon Apr 1, 2022
@yiheng-wang-nv is investigating requirements, ideas, solutions, will prepare a proposal soon.
Thanks.
[-]Automatically download / upload models from hugging face, torch hub[/-][+]Automatically download / upload models with hugging face, torch hub[/+]Nic-Ma commentedon Apr 4, 2022
Hi @yiheng-wang-nv ,
In my simple mind, I think maybe we can start from the
downloadfeatures and think about below initial ideas first: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
2.1. 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)
3.1. use
torch.hub.XXX, refer to: https://github.com/huggingface/transformers/blob/main/hubconf.py3.2. (P1) use hugging face API, refer to: https://huggingface.co/docs/hub/adding-a-library
Thanks.
4042 Hugging Face Hub integration (#6454)