Add PyTorchModelHubMixin to enable downloads on HF

[NielsRogge]

Hi @gastruc,

Thanks for this nice work and integrating with the hub already!

Fixes #1

I wrote a quick PoC to showcase that makes HF integration even easier by leveraging the PyTorchModelHubMixin class which allows to:

• track download numbers for your models (similar to models in the Transformers library)
• leverage safetensors for weights serialization
• add from_pretrained and push_to_hub automatically.

Usage is as follows:

from hubconf import AnySat

# load model
model = AnySat(...)

# equip with weights
model.load_state_dict_from_url("https://huggingface.co/g-astruc/AnySat/blob/main/models/AnySat.pth")

# push to the hub
model.push_to_hub("nielsr/my-awesome-anysat-model")

# great! We can now use it as follows (+ this increments download numbers)
model = AnySat.from_pretrained("nielsr/anysat-demo")

Would you be interested in this integration?

Kind regards,

Niels

[gastruc]

Hi,
Thank you for your contribution!
It seems that your PR also remove torch hub integration. I'd like to keep that :)

[NielsRogge]

Hi, thanks for your reply.

The model is still compatible with Torch hub, it can be loaded as follows:

import torch

AnySat = torch.hub.load('gastruc/anysat', 'anysat', pretrained=True, flash_attn=False)

Since this PR does not touch the hubconf.py file. You could theoretically add both .pt and safetensors weights in the gastruc/anysat repository.