Integration with the Hugging Face Hub

Is your feature request related to a problem? Please describe. As I train models, I would like to easily be able to share them with other people and document them well. I would also like to be able to access other trained models from the community.

Describe the solution you’d like I would like to have an integration with the Hugging Face Hub (disclaimer: I’m a member of the OS team there). I would like to be able to do model.push_to_hub("osanseviero/my_vae") and get a model directly in the Hub. Some of the benefits of sharing models through the Hub:

• versioning, commit history and diffs
• repos provide useful metadata about their tasks, languages, metrics, etc that make them discoverable
• multiple features from TensorBoard visualizations, leaderboards, and more

[clementchadebec]

Hi @osanseviero, Thank you for requesting this feature. It was added in #28 and we can now share and download models though the HuggingFace Hub! To do so you need:

• a valid HuggingFace account
• the package huggingface_hub installed in your virtual env. If not you can install it with

$ python -m pip install huggingface_hub

• to be logged in to your HuggingFace account using

$ huggingface-cli login

Uploading a model to the Hub

Any Pythae model can be easily uploaded using the method push_to_hf_hub

>>> my_vae_model.push_to_hf_hub(hf_hub_path="your_hf_username/your_hf_hub_repo")

Note: If your_hf_hub_repo already exists and is not empty, files will be overridden. In case, the repo your_hf_hub_repo does not exist, a folder having the same name will be created.

Downloading models from the Hub

Equivalently, you can download or reload any Pythae model directly from the Hub using the method load_from_hf_hub

>>> from pythae.models import AutoModel
>>> my_downloaded_vae = AutoModel.load_from_hf_hub(hf_hub_path="path_to_hf_repo")

[osanseviero]

cc @nateraw @nimaboscarino

Yes, the way we add metadata is by creating a README.md file with yaml metadata at the top. Having the small snippet you suggested would work great. huggingface_hub has a couple of utilities to make updating the metadata very very easy. For example, you can use metadata_update to just add the tag. We also have a couple of nice utilities being added to huggingface_hub for easier programmatic model card creation (https://github.com/nateraw/modelcards being merged right now), but the existing tools should work well already. Let us know if you have any feedback!