Potential bug in Qwen 2/2.5 VL Image Preprocessor

[ritwickchaudhry]

transformers/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py

Line 278 in 5c47d08

repeats = np.repeat(patches[-1][np.newaxis], temporal_patch_size - 1, axis=0)

The temporal_patch_size is used to group consecutive video frames. However, if the number of frames are not divisible, then the last frame is repeated. The current number of repetitions is temporal_patch_size - 1. While this will work for temporal_patch_size = 2 but it wouldn't work for larger patch sizes.

In my opinion, the code should be modified to:

repeats = np.repeat(patches[-1][np.newaxis], temporal_patch_size - (patches.shape[0] % temporal_patch_size), axis=0)

[zucchini-nlp]

@ritwickchaudhry correct! There was another issue with PR somewhere (#37350), probably got stale. I have forgot about it due to low priority. Would you like to open a PR for this?

LMK if you can't contribute, I can finalize and merge the existing PR later next week :)

[ritwickchaudhry]

Thanks @zucchini-nlp ! Sure, let me create a PR soon!

[anshulsc]

Hi @ritwickchaudhry and @zucchini-nlp,

I've also encountered this issue and have implemented the fix based on the discussion here.
@ritwickchaudhry, I saw you mentioned you'd be creating a PR. Are you still planning to, or would you mind if I submit one with the changes?

[ritwickchaudhry]

@anshulsc I'll be releasing the PR very soon, as I finished most of it. Thanks for the offer though!

[anshulsc]

@ritwickchaudhry great !!

[ritwickchaudhry]

Done actually! @zucchini-nlp can you please review the PR: #38076