TopolearnReal tuning runs need time budgets, early stopping of the whole study, and a head start from prior knowledge.
study.optimize stops when either limit is reached:
n_trials=100: number of trials in this call (not counting earlier ones in a resumed study).timeout=3600: seconds; the running trial finishes, no new one starts.
With neither, it runs until interrupted or study.stop() is called.
Callbacks are functions callback(study, frozen_trial) called after each trial, passed via optimize(..., callbacks=[...]). Uses include logging to your experiment tracker, saving the best model, or stopping when no improvement happened for N trials by calling study.stop(). optuna.study.MaxTrialsCallback(n, states=...) caps the total number of trials across all workers sharing a storage.
Enqueue known configurations with study.enqueue_trial({"lr": 1e-3, "depth": 6}) before optimizing. The next trial uses those values instead of sampling (missing parameters are sampled), so defaults or previous best results are evaluated first and give the sampler a strong starting point.
Other useful options:
show_progress_bar=Truefor a tqdm bar.gc_after_trial=Trueto run garbage collection between memory-heavy trials.study.add_trial(optuna.trial.create_trial(params=..., distributions=..., value=...))to import results computed outside Optuna.
Example
import optuna
optuna.logging.set_verbosity(optuna.logging.WARNING)
def objective(trial):
x = trial.suggest_float("x", -10, 10)
return (x - 2) ** 2
class StopWhenStale:
def __init__(self, patience):
self.patience = patience
def __call__(self, study, trial):
if trial.number - study.best_trial.number >= self.patience:
study.stop() # no improvement for `patience` trials
study = optuna.create_study(sampler=optuna.samplers.TPESampler(seed=0))
study.enqueue_trial({"x": 1.5}) # evaluate a known-good guess first
study.optimize(objective, n_trials=500, timeout=60, callbacks=[StopWhenStale(25)])
print(study.trials[0].params) # {'x': 1.5}
print(len(study.trials) < 500) # True: stopped early by the callbackRemember
n_trialsandtimeoutbound a run;study.stop()ends it from a callback- Callbacks receive (study, frozen_trial) after every trial
enqueue_trialevaluates chosen parameters before sampling starts
Watch out for
- Expecting timeout to interrupt a running trial; it only prevents new trials from starting.