added zero shot image classification

[dulayjm]

👋 Hi, first-time potential contributor here! I noticed no method for zero-shot-image-classification in the inference api, and I wrote an implementation to call for that task.

🤗 Solution:

• Add post response in _client.py to call a zero-shot image classification model (such as CLIP).
• Added tests and relevant documentation.

TODO:

Test other models

[Wauplin]

Hi @dulayjm! Thanks for opening a PR for this task 🎉 Not all tasks are indeed implemented as the client is quite new and we want other tasks to be more community-lead so I'm glad to see your contribution 🤗

I reviewed your PR and made some comments to it. I hope everything's clear but don't hesitate to ping me if not. In general, it would be good to review the inputs/output of the method and adapt the corresponding docs/tests :)

[Wauplin]

This file must be generated automatically (here the format is not correct). To do so, you must have pytest-vcr installed + run pytest tests/test_inference_client.py -k "zero" --vcr-record=all to record the HTTP call. Don't worry if it doesn't work for you, I can take of it :) Just be careful of not committing your HF token in the yaml file.

[dulayjm]

This was extremely helpful! I could not find the command to generate this properly. With the typing changes, it passes all tests!

[dulayjm]

@Wauplin Thank you for the suggested changes!! These were very helpful! In response, I implemented the following:

• Addressed a typo in the docs
• Change the input type to List[str] to enforce a precise expected input type for labels
• Added a raise to ValueError if the length of label responses is less than 2.
• Changed the output response type to reflect the OpenCLIP model List[ClassificationOutput]
• Created the VCR testing file
• Updated a better test in test_inference_client.py to address the typing and expected outputs for the model and task.

[Wauplin]

Really nice! I made a quick review of the code and left a few minor comments. I'll run some tests and let you know for a final go!

[Wauplin]

Perfect! Great job here @dulayjm 👏 I'll merge it as soon as the CI is green 🟢

Would you mind sharing for what project you're interested in using this client? I'd be glad to know more :)

[dulayjm]

@Wauplin Thanks for the review! Happy to see this get merged for HF 🤗

I'm using this for a project within my company to quickly deploy CLIP (and a variant of it, fashion-clip to quickly and accurately label items of clothing that our users at Capsule (YC 22 - coming soon!!) can upload to the platform!

[Wauplin]

Nice!