Daniel Rasmussen

Looks like that change was un-reverted in https://github.com/keras-team/keras/pull/20051. However, after doing some quick testing in Keras 3.5, the behaviour is different but still incorrect. Running the same script from here...

Thanks for looking into this @james77777778 / @fchollet, but I don't think this was fixed in #20541. When I run the test script from https://github.com/keras-team/keras/issues/19891#issuecomment-2183215799 on the latest `master` commit...

The updated test script you need to run is here https://github.com/keras-team/keras/issues/19891#issuecomment-2183215799, it doesn't involve evaluate.

Can confirm that this was fixed in #20609, thanks very much @james77777778!

The shape of the mask is correct in this example (according to https://github.com/keras-team/keras/issues/19818#issuecomment-2156142266), so validation wouldn't help in this case.

If you install the current development version from the main branch it should be compatible with tensorflow 2.16. We'll do a release with those updates soon.