[GENERAL SUPPORT]: Saving + loading experiment including trial data

[dthillaithevan]

Question

Hi,

What is the correct way to store and load an experiment and current trial data? My goal is to enable the ability to restart an optimization from any iteration using previously stored data.

The method below using to_json_snapshot() and load_from_json_file does not seem to work. Attempting to load the json data into another client and calling get_best_trial() results in an empty data error.

Thanks!

Please provide any relevant code snippet if applicable.

# Setup experiment
...

# Run N iterations of BO
for i in range(N_iter):
    params, it = ax_client.suggest_parameters()
    objective_value = objective(params)
    ax_client.update_optimizer(it, objective_value)
    
    # Save trial to json
    
    json_data = ax_client.to_json_snapshot()

    # Save data in each iteration
    with open("./exp.json", "w") as f: 
        json.dump(json_data, f)
        

# Load experiment + data from json
ax_client_2 = AxClient()
ax_client_2.load_from_json_file("./exp.json")

# This results in: Encountered error while trying to identify the best point: Cannot identify best point if experiment contains no data.
ax_client_2.get_best_trial()

Code of Conduct

• I agree to follow this Ax's Code of Conduct

[lena-kashtelyan]

I'm not sure where the suggest_parameters and update_optimizer calls in your example come from –– AxClient does not expose such methods (where did you get those, just for my knowledge?).

I think if you use AxClient.get_next_trial/AxClient.gen_next_trials and AxClient.complete_trial (see tutorial: https://ax.dev/docs/tutorials/gpei_hartmann_service), JSON storage should work.