Keras Issue in reduce_shape in keras/src/ops/operation_utils.py

keras.ops.any and keras.ops.all seem to be the only operators in latest docs that specifically say axis may be negative. However, looking through the code, both these ops and other reduction ops (like keras.ops.amax, for example) use the same utility function to calculate the output shape: reduce_shape in keras/src/ops/operation_utils.py.

I think this function does not handle lists/tuples of negative indexes correctly. The following example shows that, running in keras 3.11.3

import keras
x = keras.Input((3, None, None, 1280))

max_x = keras.ops.amax(x, axis = (2, 3))
assert max_x.shape == (None, 3, 1280)  # passes

max_x = keras.ops.amax(x, axis = (-2, -3))
assert max_x.shape == (None, 3, 1280)  # does not pass, despite being equivalent

I believe this can be solved in the reduce_shape function by first doing something like axis = [ax % len(shape) for ax in axis]. I can submit a pull request for this later on (I never contributed to keras, so it could be a nice initiation)

Let me know if there's important info that's missing, I did not find a template for issues that I should follow.

Comment From: pctablet505

looks like issue is fixed