Inconsistent manner of the metric `SpecificityAtSensitivity` among different backends about keras HOT 5 OPEN

Hi @Star9daisy ,

I have replicated the reported behaviour and attached gist here.Need to dig more and will come back.

from keras.

Possibly related to #19433

from keras.

Thank you @Star9daisy! Would you be able to submit a fix for this? Otherwise we're leaving this open for contributions from the community.

from keras.

Hi @grasskin, it looks kind of difficult for me. But I'd like to give it try!

from keras.

Go ahead @Star9daisy! The nonzero error is likely related to #19407 which we just closed with a fix. Feel free to update this thread with concerns/questions. One thing I would strongly recommend is disabling XLA compilation to make debugging easier. Specifically tf.config.run_functions_eagerly(False) for tensorflow and jax.config.update("jax_disable_jit", True) for JAX.

You will also likely benefit from disabling Keras traceback filtering keras.config.disable_traceback_filtering() to get full error tracebacks.

from keras.