search

weaviate / i2v-pytorch-models

Inference containers for the Weaviate `img2vec-pytorch` module
BSD 3-Clause "New" or "Revised" License
1 stars 4 forks source link

Feature Request: Use Hugging Face Model Hub instead of Torch Hub to load modules #7

Open hsm207 opened 1 year ago

hsm207 commented 1 year ago

The current implementation use torch hub to load the resnet model.

We should use the transformers library instead because it integrates well with hugging face's model hub which makes it easy for users to use their own models for vectorization.

Resnet 50 is also on model hub.

CShorten commented 1 year ago

How about, in addition to, rather than instead of -- we could follow the same naming convention of img2vec-transformers & img2vec-huggingface, while keeping img2vec-pytorch in place. I still imagine people will want to optimize their own computer vision models native to pytorch or keras/TF/jax at that.

hsm207 commented 1 year ago

I still imagine people will want to optimize their own computer vision models native to pytorch or keras/TF/jax at that.

This is a good point. We will have to weigh this against the pros and cons of maintaining 2 almost functionally identical modules.

Can we ask our community to vote on this?