I used TPESampler and set it as follows while optimizing with optuna: sampler=optuna.samplers.TPESampler(multivariate=True, n_startup_trials=10, seed=None). But in the 10 startup_trials process, it always takes exactly the same parameters for the trial. I conducted additional experiments with RandomSampler and found that it also has the same situation, always collecting exactly the same parameters. Because I used Flax, which does not support multiple processes, I had to encapsulate the processes. If anyone could help me, I would be extremely grateful. My code as below. def multiprocessing_objective_fn(args, trial): queue = multiprocessing.Queue() p = multiprocessing.Process(target=train_agent, args=(args, trial, queue)) p.start() p.join() result = queue.get() return result if name == "main": from optuna.storages import RDBStorage from sqlalchemy import create_engine import socket args = get_args() print(args) # Step 2: Create the engine with the specified timeout engine = create_engine("sqlite:///optuna_database_250710_test5" + ".db", connect_args={'timeout': 500}) # Step 3: Use this engine to create the Optuna storage storage = RDBStorage("sqlite:///optuna_database_250710_test5" + ".db") study = optuna.create_study( direction="maximize", storage=storage, load_if_exists=True, study_name=args.exp_name + '__' + args.env_id + '__' + args.run_name + '__' + args.actor, pruner=optuna.pruners.HyperbandPruner() ) objective_fn = functools.partial(multiprocessing_objective_fn, args) #objective_fn = functools.partial(train_agent, args) # wandb only works for n_jobs = 1 ! See README.md for more infos about that study.optimize(objective_fn, n_trials=args.n_trials, n_jobs=1) Continue reading...
