Add Hugging Face integration

[NielsRogge]

Hey,

thanks for this great model!

I though it'd be nice to leverage the PyTorchModelHubMixin class, which allows to add from_pretrained and push_to_hub capabilities to any custom nn.Module, which means you'll get download metrics for your models, people can easily discover these models (if we tag them with "3D vision", etc.)

Now you can do:

from dust3r.model import AsymmetricCroCo3DStereo

model = AsymmetricCroCo3DStereo.from_pretrained("nielsr/DUSt3R_ViTLarge_BaseDecoder_512_dpt")

The model is here: https://huggingface.co/nielsr/DUSt3R_ViTLarge_BaseDecoder_512_dpt_bis.

Here's a notebook to showcase its usage: https://colab.research.google.com/drive/1CQu_sAWhGyuKcuXCqOB4lunqXgOXCeAW?usp=sharing.

Of course, we could move the checkpoints to the NAVER organization on the hub rather than my personal username (nielsr).

Kind regards,

Niels ML @ HF

[NielsRogge]

Gentle ping @yocabon

[NielsRogge]

Hi @yocabon I saw you pushed a commit. As can be seen, download metrics are working (shown on the right):

If you could push the Dust3R checkpoints under your username (or the organization name of NAVER): https://huggingface.co/naver, I can update the README in this PR to replace my username "nielsr" by the new name.

Would you be able to push them? I'm happy to help.

Basically now people don't need to manually download a checkpoint first, it will download automatically from the hub + increment the download counts

[yocabon]

Hi, thanks for the suggestions. Because I'd like to make the integration optional and more importantly to keep being able to load local checkpoints; I made it a little more complicated in the branch. I managed to upload the checkpoints to the naver organization without any issue (for now it's still in private). I'll discuss this with the others and maybe merge later if they're okay with it.

Please tell me if you also have remarks.

[yocabon]

Done with #78

[NielsRogge]

Thanks @yocabon, that works for me. The integration looks nice! cc @wauplin for visibility

[NielsRogge]

Hi @yocabon looks like the HF integration is a great success, the model has been downloaded more than 200k times last 30 days! 🚀

I opened a PR to add the "mixin" tag so we can track the downloads: https://huggingface.co/naver/DUSt3R_ViTLarge_BaseDecoder_512_dpt/discussions/1. This way we can see which models which are using the mixin have the most amount of downloads.

Thanks!