Recent PRs have highlighted increasing issues with mixing keras and tensorflow - things like random seeds for reproducible testing, callbacks, and minor API changes and deprecation warning. While keras is still maintained, compatibility between versions is ill-documented and the general assumption appears to be that ppl should use tf.keras - in fact, keras.io now exclusively documents the latter.
Recent PRs have highlighted increasing issues with mixing keras and tensorflow - things like random seeds for reproducible testing, callbacks, and minor API changes and deprecation warning. While keras is still maintained, compatibility between versions is ill-documented and the general assumption appears to be that ppl should use tf.keras - in fact, keras.io now exclusively documents the latter.