Question about intended behavior of is_continuous_axis for decreasing axes

[MalyalaKarthik66]

Hi Keras team,

While contributing and reading through keras/src/utils/python_utils.py, I noticed the implementation of is_continuous_axis.

Currently, the function defines a negative_order_flag, but the loop checks the same condition as the positive case (axis[i + 1] - axis[i] != 1). From local testing, decreasing contiguous sequences (e.g., [3, 2, 1, 0]) currently return False.

I wanted to check the intended behavior here:

• Was this function meant to treat decreasing contiguous axes as continuous (as the negative_order_flag name might suggest)?
• If the condition were changed to check != -1, could that introduce issues for backend kernel selection paths (for example in nn.py for normalization layers)?

I have a small patch and unit tests prepared, but wanted to confirm the intended semantics before opening a PR.

Thanks!

[Sameekshagithub]

Hi,

I looked at the behavior you described. It seems that the current implementation only treats increasing consecutive axes as continuous.

The presence of negative_order_flag suggests that decreasing sequences may have been considered, but if the condition still checks:

axis[i + 1] - axis[i] != 1

then a sequence like [3, 2, 1, 0] will not satisfy the condition because the difference is -1.

I think confirming the intended semantics would be important before changing this, because is_continuous_axis may be used in backend-specific paths where the ordering of axes could affect behavior.

A good approach would be:

1. Check existing unit tests for expected behavior with decreasing axes.
2. Search for usages of is_continuous_axis to see whether reverse ordering is supported anywhere.
3. If decreasing axes are intended to be valid, adding tests for cases like [3,2,1,0] and updating the condition to handle -1 would make the behavior explicit.

Thanks for raising this. Clarifying the expected behavior through tests would help avoid unintended changes.