Upcast rotated box transforms #9175
Open
Conversation
This file contains hidden or 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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/vision/9175
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 New FailuresAs of commit 0ff4716 with merge base b818d32 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
NicolasHug
approved these changes
Aug 12, 2025
There was a problem hiding this comment.
Thanks a lot for noticing the issue and for the fix @AntoineSimoulin . I left a comment below but I'll approve to unblock
| @@ -451,6 +451,12 @@ def _parallelogram_to_bounding_boxes(parallelogram: torch.Tensor) -> torch.Tenso | |||
| torch.Tensor: Tensor of same shape as input containing the rectangle coordinates. | |||
| The output maintains the same dtype as the input. | |||
| """ | |||
| dtype = parallelogram.dtype | |||
| acceptable_dtypes = [torch.float32] | |||
There was a problem hiding this comment.
I think we should consider float64 to also be an acceptable dtype? Otherwise we risk converting a float64 into a float32, which may be unnecessary.
Suggested change
| acceptable_dtypes = [torch.float32] | |
| acceptable_dtypes = [torch.float32, torch.float64] |
| @@ -278,9 +251,11 @@ def _xyxyxyxy_to_xywhr(xyxyxyxy: torch.Tensor, inplace: bool) -> torch.Tensor: | |||
| xyxyxyxy = xyxyxyxy.clone() | |||
|
|
|||
| dtype = xyxyxyxy.dtype | |||
| need_cast = not xyxyxyxy.is_floating_point() | |||
| acceptable_dtypes = [torch.float32] # Ensure consistency between CPU and GPU. | |||
There was a problem hiding this comment.
Suggested change
| acceptable_dtypes = [torch.float32] # Ensure consistency between CPU and GPU. | |
| acceptable_dtypes = [torch.float32, torch.float64] # Ensure consistency between CPU and GPU. |
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.
Summary
This pull request addresses an issue where performing certain operations on rotated bounding boxes in float16 format could lead to NaN values due to overflow.
Issue Details
The functions _xyxyxyxy_to_xywhr and _parallelogram_to_bounding_boxes involve squaring operations on bounding box coordinates. When using float16 precision, boxes with absolute coordinate values above 256 can result in overflow, as the square of 257 ($257^2 = 66,049$ ) exceeds the maximum representable value for float16 ($65,504 = (2−2^{−10}) \times 2^{15}$ ). This leads to unreliable results and NaN values in subsequent computations.
Solution
To address this issue, this PR ensures that the floating point data is upcast to at least float32 for these operations. This increases the maximum representable value, allowing support for box coordinates up to$\sqrt{2^{31} - 1} \approx 46,341$ , which should be sufficient for all common computer vision applications.
Additional Improvements
Simplified code related to rotated bounding box conversion by removing unnecessary integer-to-float conversions.
Rotated Bounding Boxes should not be constructed with integer dtype, so these conversions were redundant.