Fixing class embedding selection in owl-vit

[orrzohar]

What does this PR do?

Fixes # For OWL-ViT image-guided object detection, there is a mistake in selecting the best embedding (the most distinct one with high IoU). Specifically;
selected_inds is a [num_inds, 1] dimensional tensor, where the indexes indicate which queries had a high IoU with target object bbox. But, as selected_inds[0] was selected, only the first of all the possible queries is selected. Specifically, instead of selected_embeddings being a [num_inds, D] dimensional tensor, it is a [1, D] dimensional tensor. This led ultimately to the first query always being selected, not the most unique one as required.

An error is not raised. To see this is the case, just add a print statement of 'torch.argmin(mean_sim)' here:

transformers/src/transformers/models/owlvit/modeling_owlvit.py

Line 1505 in 01734db

best_box_ind = selected_inds[torch.argmin(mean_sim)]

& you will see it is always 0.

• transformers version: 4.28.1
• Platform: Linux-5.4.0-109-generic-x86_64-with-glibc2.31
• Python version: 3.11.3
• Huggingface_hub version: 0.14.1
• Safetensors version: not installed
• PyTorch version (GPU?): 2.0.0+cu117 (True)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using GPU in script?: Yes

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@sgugger
@NielsRogge
@alaradirik
@amyeroberts

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[alaradirik]

@orrzohar thank you for opening the PR! I'll take double check the original code and the forward pass shortly.

[alaradirik]

Thank you for catching this and opening the PR! I double checked the corresponding part in the original repo and the change indeed fixes the issue (as documented in issue #21206).

Adding @amyeroberts for the final approval.

[MaslikovEgor]

After this fix I have problem with predictions.

I Use Colab demo for owl-vit
And on example photo with cats I have only one prediction. Looks suspicious and not the same as original repo

@alaradirik

[MaslikovEgor]

Hi @MaslikovEgor, Hub demos are deployed once and not updated unless triggered. I'm rebooting the demo to reflect the changes.

Nice to meet you!

Yeah, I understand this. But in google colab demo we install fresh transformers from source:

!pip install git+https://github.com/huggingface/transformers.git

So this is the problem with this changes

@alaradirik

[orrzohar]

I found when evaluating COCO that mAP@0.5 increases from 6 to 37. This is still below the expected 44+, but closer to the reported/expected performance.

I am still trying to figure out why.