Decoupling weights downloading from loading about curated-transformers HOT 6 CLOSED

Thanks for the feature request! This sounds like a very useful feature and would be simple to add. We'll add it to our todo list.

from curated-transformers.

If there's an easy way to also support loading from a byte string, I'd see that as desirable too 🙂

I get that it might not be so easy though, depending on the way PyTorch does things.

from curated-transformers.

If there's an easy way to also support loading from a byte string, I'd see that as desirable too 🙂

I get that it might not be so easy though, depending on the way PyTorch does things.

We could definitely support that. But we still need to hash out how we want to serialize the model hyperparameters (outside TorchScript/ONNX/etc.). What we had in mind now is generating configuration json that it compatible with HF transformers. We could then serialize the config + state dict to a byte string.

from curated-transformers.

@bilelomrani1 the main branch now supports this, so if you want to give it a spin:

https://curated-transformers.readthedocs.io/en/latest/encoders.html#curated_transformers.models.AutoEncoder.download_to_cache

from curated-transformers.

Thank you @shadeMe @danieldk this works exactly as intended! I leave the issue open if you want to track the status for loading from a byte string here but feel free to close, as this solves my issue.

from curated-transformers.

Thanks for the suggestion and for testing 👍.

from curated-transformers.