diffquant/optimize.py at main · YuriyKolesnikov/diffquant · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
# optimize.py
"""
DiffQuant — Optuna hyperparameter search entry point.

Runs a full train + val evaluation for each trial.
Optimises for best walk-forward val Sharpe.

Usage:
    python optimize.py --config configs/experiments/itransformer_hybrid.py --trials 50
    python optimize.py --config configs/experiments/itransformer_hybrid.py \\
                       --trials 100 --jobs 1 \\
                       --storage sqlite:///optuna_diffquant.db

Notes:
    - Each trial runs cfg.training.num_epochs epochs.
      Use a reduced epoch count in the config for search
      (e.g. 50–100 epochs), then retrain the best config to convergence.
    - Parallel jobs (--jobs > 1) require a persistent storage backend.
    - The best hyperparameters are printed and saved to
      output/<experiment>/optuna_best.json at the end.
"""

import argparse
import json
import logging
import sys
from pathlib import Path

logging.basicConfig(
    level   = logging.INFO,
    format  = "%(asctime)s | %(message)s",
    datefmt = "%Y-%m-%d %H:%M:%S",
    handlers = [logging.StreamHandler(sys.stdout)],
)
log = logging.getLogger(__name__)


def main() -> None:
    parser = argparse.ArgumentParser(description="DiffQuant hyperparameter search")
    parser.add_argument("--config",   "-c", required=True)
    parser.add_argument("--trials",   "-n", type=int, default=50)
    parser.add_argument("--jobs",     "-j", type=int, default=1)
    parser.add_argument("--storage",  "-s", default=None,
                        help="Optuna storage URL. Default: in-memory (not persistent).")
    args = parser.parse_args()

    try:
        import optuna
    except ImportError:
        log.error("optuna not installed. Run: pip install optuna")
        sys.exit(1)

    optuna.logging.set_verbosity(optuna.logging.WARNING)

    from configs.experiments import load_config

    base_cfg = load_config(args.config)

    def objective(trial: optuna.Trial) -> float:
        """
        One trial = one full training run with sampled hyperparameters.
        Returns val Sharpe (annualised, from walk-forward evaluation).
        Returning -inf on failure allows Optuna to continue gracefully.
        """
        import copy
        cfg = copy.deepcopy(base_cfg)

        # ── Search space ──────────────────────────────────────────────────────
        # Keep the space tight — only parameters with high expected impact.
        cfg.training.lr            = trial.suggest_float("lr",            1e-5, 5e-4, log=True)
        cfg.training.weight_decay  = trial.suggest_float("weight_decay",  1e-6, 1e-3, log=True)
        cfg.training.gradient_clip = trial.suggest_float("gradient_clip", 0.3,  2.0)

        cfg.loss.lambda_turnover = trial.suggest_float("lambda_turnover", 0.001, 0.1,  log=True)
        cfg.loss.lambda_drawdown = trial.suggest_float("lambda_drawdown", 0.001, 0.05, log=True)
        cfg.loss.lambda_terminal = trial.suggest_float("lambda_terminal", 0.0,   0.01)

        cfg.policy.tau_gate      = trial.suggest_float("tau_gate",      0.3, 3.0)
        cfg.policy.tau_direction = trial.suggest_float("tau_direction",  0.3, 3.0)

        # Unique output dir per trial to avoid checkpoint collisions.
        cfg.paths.output_dir = str(
            Path(base_cfg.paths.output_dir) / "optuna" / f"trial_{trial.number}"
        )

        try:
            import torch
            from data.pipeline    import load_or_build
            from data.dataset     import TradingDataset
            from training.trainer import DiffTrainer

            splits   = load_or_build(base_cfg.paths.source_data, cfg, base_cfg.paths.cache_dir)
            train_ds = TradingDataset(splits["train"])
            trainer  = DiffTrainer(cfg)
            trainer.train(train_dataset=train_ds, raw_val=splits["val"])

            return trainer.best_val_sharpe

        except Exception as exc:
            log.warning("Trial %d failed: %s", trial.number, exc)
            return float("-inf")

    # ── Create study ──────────────────────────────────────────────────────────
    study = optuna.create_study(
        study_name    = base_cfg.experiment_name,
        direction     = "maximize",
        storage       = args.storage,
        load_if_exists= True,
        sampler       = optuna.samplers.TPESampler(seed=base_cfg.seed),
        # pruner removed: requires trial.report() calls inside training loop
        # pruner        = optuna.pruners.MedianPruner(n_warmup_steps=5),
    )

    study.optimize(
        objective,
        n_trials  = args.trials,
        n_jobs    = args.jobs,
        show_progress_bar = True,
    )

    # ── Report ────────────────────────────────────────────────────────────────
    best = study.best_trial
    log.info("=" * 52)
    log.info("  OPTUNA SEARCH COMPLETE")
    log.info("  Best trial   : #%d", best.number)
    log.info("  Best Sharpe  : %+.4f", best.value)
    log.info("  Best params  :")
    for k, v in best.params.items():
        log.info("    %-25s = %s", k, v)
    log.info("=" * 52)

    # Save best params
    out_path = Path(base_cfg.paths.output_dir) / "optuna_best.json"
    out_path.parent.mkdir(parents=True, exist_ok=True)
    with open(out_path, "w") as f:
        json.dump({
            "best_trial":  best.number,
            "best_sharpe": best.value,
            "params":      best.params,
        }, f, indent=2)
    log.info("Best params saved → %s", out_path)


if __name__ == "__main__":
    main()