Possibly unintended regression (though maybe intended?) about keras

[fchollet]

When you do input_two[..., None] (where input_two is a KerasTensor) you are hitting keras/src/ops/numpy.py:GetItem.compute_output_spec(), which supports Ellipsis.

This is why your model can be built and trained. You'd see an error at this line otherwise.

However the framework doesn't know how to serialize Ellipsis. I think we can support it in the general case, by adding a case for it in serialize_keras_object/deserialize_keras_object (in keras/src/saving/serialization_lib.py). Similar to what we do for __slice__. If you're able, please open a PR.

[LukeWood]

Sounds good - I'll give this a shot and see if I can add some tests + add support later today.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No