Optuna & Ray Tune

What it is

Optuna for clean Bayesian search inside one process; Ray Tune for distributed trials across a cluster. Both implement Hyperband, ASHA and population-based training.

How Vaaani uses it

• Squeezing the last 3% accuracy from a tuned model
• Distributed trials across spot instances
• Multi-objective optimization (accuracy + latency)
• Pruning bad trials early to save compute budget

Why it makes the cut

Manually tuning hyperparameters wastes budget. Optuna's pruner finds the answer 5x faster and you go to lunch.

Sample code

import optuna

def objective(trial):
    lr = trial.suggest_loguniform("lr", 1e-5, 1e-1)
    n  = trial.suggest_int("n_estimators", 50, 500)
    return train_eval(lr, n)

study = optuna.create_study(direction="maximize")
study.optimize(objective, n_trials=100)