Upsampling2d Skips one sample at index 8388608 (2^23)

[jrasko]

When using UpSampling2D in mode "nearest", one index is skipped at exactly 8388608 (2^23). This results in a shift of all following data. See the image below:

At the end, the shape is fixed by just padding the last data:

Example Code:

import numpy as np
from keras import layers

a = np.zeros(2 ** 22 + 2)
a[1::2] = 1
up = layers.UpSampling2D(size=(2, 1))
b = up(a[None, :, None, None])
print(a[-6:].flatten())
# output: [0. 1. 0. 1. 0. 1.]
print(b[:, -6:, :, :].numpy().flatten())
# output: [1. 1. 0. 1. 1. 1.]

[jrasko]

Seems like this occurs in tensorflow, but not when using the newest keras version. Therefore, it seems like this is already fixed

[mehtamansi29]

Hi @jrasko -

This issue is only on keras 2.15.0 but this issue doesn't reproduce in keras 3.3 and keras 3.4 latest version.