Keras Inconsistent execution results by the PyTorch backend

When using keras.layers.MaxPooling2D with PyTorch backend, there is an inconsistent execution result between static inference shape and dynamic results.

import os
import re
import torch
import numpy as np
os.environ['KERAS_BACKEND']='torch'
import keras

layer = keras.layers.MaxPooling2D(
    pool_size=[ 2, 3 ],
    strides=[ 3, 3 ],
    padding="same",
    data_format="channels_first",
    trainable=True,
    autocast=True,
)

result_static = layer.compute_output_shape([1, 5, 5, 4])

result_dynamic = layer(
    inputs=np.random.rand(*[1, 5, 5, 4]),
)

The version is keras 3.5.0 with PyTorch 2.4.0 And I got the results below

result_static: 
(1, 5, 2, 2)

result_dynamic.shape:
torch.Size([1, 5, 2, 1])

Comment From: sachinprasadhs

Thanks for reporting the issue, on investigation it was found out that this happens with all the layers in torch which uses padding="same". It was kind of edge case scenario with only certain combination like you provided would catch the error. Created a PR to fix the same.

Comment From: dhantule

Hi @MilkFiish, thanks for reporting this.

This issues seems to be fixed in this PR, I have tested your code with the keras 3.10.0 and it's working fine in this gist.

Comment From: github-actions[bot]

This issue is stale because it has been open for 14 days with no activity. It will be closed if no further activity occurs. Thank you.