Does distributed Keras tuner save a "oracle.json" to allow for tuner reloading with trials?

[hmf]

I am a little confused about reloading a tuner to access the trials. I made some initial experiments and during one of these session I incorrectly used the wrong library (imported the keras library, not from TF). In this case I see that the Keras tuner saved a JSON file per slave. It also saved an oracle.json file for the chief. When I reload the tuner I can access the trials with no problem because the tuner picks up the oracle.json file.

However, when I use the "correct" import, I see no oracle.json file. I cannot load the data irrespective of the KERASTUNER_TUNER_ID I use. This is what I have in my code:

    os.environ['KERASTUNER_TUNER_ID'] = 'oracle'

So the question is, does the Tuner save a oracle.json file? If not, how do I get this file. This code seems to indicate that I need not set parameters explicitly.

I am also using overwrite = Falseand setting the directory and project_name parameters to the same values used during tuning. I also delete the directory folder when doing a new tuning session.

I have searched this issue but nobody seems to have complained.

TIA

EDIT: one small detail I neglected to mention. When the tuning finishes, the chief process dos not terminate. I am forced to kill it.

[hmf]

Finally figured this out. I had created (constructed) 2 tuners in the same process. I then used the second one to do the search. It seems like the path to the oracle JSON file that is set in the 1st constructor is not used on the second one. The other members/properties are updated correctly. The result is the output of the second tuner's oracle was going to the folder set up in the first tuner (other data was set correctly).

Does not seem intuitive. I assume that searching on the tuner will always use that tuner's properties only.