| """Experiment runner. Usage: python -m experiments.run <model> <config_json> |
| |
| Prints the final line as: |
| METRIC: {json} |
| |
| so the outer loop can grep a single line instead of parsing full stdout. |
| """ |
| from __future__ import annotations |
|
|
| import json |
| import sys |
| import time |
|
|
| from experiments import eval as eev |
| from experiments import models |
|
|
|
|
| def main() -> None: |
| exp_name = sys.argv[1] |
| cfg = json.loads(sys.argv[2]) if len(sys.argv) > 2 else {} |
|
|
| model_name = cfg.get("model", exp_name) |
| context = cfg.get("context", 168) |
| horizon = cfg.get("horizon", 24) |
| on = cfg.get("on", "test") |
| batch_size = cfg.get("batch_size", 128) |
|
|
| bas = cfg.get("bas", ["PJM"]) |
| split = eev.load_split(bas) |
| print(f"[data] bas={list(split.bas)} denom_mae(macro)={split.denom_mae:.2f}", flush=True) |
|
|
| t0 = time.time() |
| forecaster = models.get(model_name, cfg) |
| load_s = time.time() - t0 |
|
|
| bootstrap = cfg.get("bootstrap", 0) |
| seed = cfg.get("seed", 0) |
| t0 = time.time() |
| m = eev.rolling_eval(forecaster, split, on=on, context=context, horizon=horizon, |
| batch_size=batch_size, bootstrap=bootstrap, seed=seed) |
| eval_s = time.time() - t0 |
|
|
| out = { |
| "exp": exp_name, "model": model_name, "on": on, |
| "bas": list(split.bas), "context": context, "horizon": horizon, |
| "load_s": round(load_s, 2), "eval_s": round(eval_s, 2), |
| **{k: round(v, 4) if isinstance(v, (float, int)) and not isinstance(v, bool) else v |
| for k, v in m.items() if k != "per_ba"}, |
| } |
| if "per_ba" in m: |
| out["per_ba"] = {ba: {k: round(v, 4) if isinstance(v, float) else v |
| for k, v in stats.items()} |
| for ba, stats in m["per_ba"].items()} |
| print("METRIC:", json.dumps(out), flush=True) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
