Zero-Shot Image Classification
OpenCLIP
Safetensors

Deployment of Model with vector Database

#2
by kreabs - opened

Helloo

i have two rather pragmatic questions:

  1. Is there a leaderboard fo text/image retrival in german and how does this model perform in german in your opinion?
    -> my tests were quite promising here but i dont have a good comparison/metric
  2. How to use this model with a vector database (like milvus)?
    -> Can i just get the vectors like in the code sample:

with torch.no_grad(), torch.autocast("cuda"):
image_features = model.encode_image(image)
text_features = model.encode_text(text)
image_features /= image_features.norm(dim=-1, keepdim=True)
text_features /= text_features.norm(dim=-1, keepdim=True)

and perfrom a cosine similarity on ethem to get the distance in milvus?

I have tried this and the results are not so good. I think that either the german language or milvus is the problem.

Thanks in advance for the answer and continue the great work

Kraebz

Hi Kraebz,

there are benchmarks at https://github.com/LAION-AI/CLIP_benchmark that support german and where you can try it on, please have a look there.
This model itself is only trained on english though.
You can try NLLB CLIP (https://arxiv.org/abs/2309.01859) which is multilingual.
See also https://github.com/mlfoundations/open_clip/blob/main/docs/openclip_multilingual_retrieval_results.csv for results
of different models on multilingual benchmarks.

Best,
Mehdi

Sign up or log in to comment