output_padding argument in Conv1DTranspose

[olive380]

Surprisingly, with TensorFlow 2.16, it is no more possible to set output_padding=0 when using Conv1DTranspose. This feature was allowed in previous versions of TensorFlow (e.g. 2.15) and is still necessary. It appears that this issue is related to the way Keras 3.0 is now managing the output_padding argument.

As reminder, in previous versions of Keras, the behaviour was totally different if output_padding was omitted or set to zero.
If you are not aware of this, please have a look at the deconv_output_length method defined in /tensorflow/python/keras/utils/conv_utils.py available here.

To reproduce the issue, please use the following python script models.zip.