Fix LLaVA-NeXT handling of non-square images #2097
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Before this change, the number of reserved image tokens was not the same as the number of images. Fixes #2029. While at it, also remove all the image token handling duplication in `prepare_input`.
We could get shape mismatches with non-square images, resulting in an exception that crashed the backend. When post-processing an image, features corresponding to padding are removed when padding was needed. This is also reflected in the calculation of the number of image tokens to get the correct number of slots. However, there was a mismatch between the post-processing and the slot calculation. The image post-processing could exclude fewer padding features due to rounding. This change updates the image token calculation to correspond to the image postprocessing. Fixes #1777. While investigating this, I found another issue where the upstream code contains a bug that swaps the height and width dimensions after computing the image grid shape. Since the models were also trained with this bug, we should reproduce the same bug to ensure that we are generating the same features.
4 tasks
Narsil
reviewed
Jun 24, 2024
| @@ -39,15 +39,14 @@ def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size): | |||
| return height // patch_size, width // patch_size | |||
|
|
|||
|
|
|||
| def image_text_replacement(image_input, config, image_id) -> str: | |||
| def image_text_replacement(processor, image_input, config, image_id) -> str: | |||
| if config.model_type == "idefics2": | |||
| # TODO technically depends on image splitting which is not implemented. | |||
There was a problem hiding this comment.
Suggested change
| # TODO technically depends on image splitting which is not implemented. |
Narsil
reviewed
Jun 24, 2024
| image_seq_len = 64 | ||
| image_str = f"<fake_token_around_image>{'<image>' * image_seq_len}<fake_token_around_image>" | ||
| if processor.image_processor.do_image_splitting: | ||
| image_str *= 5 |
There was a problem hiding this comment.
We still need that.
Instead of doing it that inefficiently maybe we just conditionally prefix/postfix (Also to make the intent clearer). Wdyt ?
Narsil
reviewed
Jun 24, 2024
|
Closing, was merged as part of #2080. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
What does this PR do?
We could get shape mismatches with non-square images, resulting in an
exception that crashed the backend.
When post-processing an image, features corresponding to padding are
removed when padding was needed. This is also reflected in the calculation
of the number of image tokens to get the correct number of slots.
However, there was a mismatch between the post-processing and the slot
calculation. The image post-processing could exclude fewer padding features
due to rounding. This change updates the image token calculation to
correspond to the image postprocessing.
Fixes #1777.
While investigating this, I found another issue where the upstream code
contains a bug that swaps the height and width dimensions after computing
the image grid shape. Since the models were also trained with this bug,
we should reproduce the same bug to ensure that we are generating the
same features.
Draft: needs to be rebased after #2080.
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
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.