"""Reproducible physical-unit benchmark on independently sampled held-out windows. Usage: python -m tfm.evaluation.benchmark --data DATA --checkpoint RUN/checkpoint.pt --output REPORT --device cpu --samples 512 --counterfactual-samples 12 Repeated --checkpoint arguments combine ensemble predictive distributions. """ from __future__ import annotations import argparse import copy import hashlib import json import resource import sys import time from pathlib import Path import numpy as np import torch from torch.utils.data import DataLoader from tfm.data import FIELD_INDEX, FIELD_SPECS, ROLE_IDS, Normalization, WindowDataset from tfm.data.schema import GENERATOR_VERSION, SCHEMAS from tfm.training.train import ( SCHEMA_VERSION, load_checkpoint, load_manifest_episodes, resolve_device, to_device, ) from .baselines import CatBoostBaseline, HistoryLinearBaseline, PersistenceBaseline from .calibration import fit_calibration, save_calibration from .counterfactual import COST_DEFINITIONS, action_effect_metrics, branch_rollouts, family_cost from .experiments import default_experiments, run_experiments, stress_observations from .metrics import metrics_by_group def sha256(path): return hashlib.sha256(Path(path).read_bytes()).hexdigest() def _ids(episodes): return sorted({str(e.system_id) for e in episodes}) class Ensemble: def __init__(self, models): self.models = models def __call__(self, batch): predictions = [model(batch) for model in self.models] means = torch.stack([x[0] for x in predictions]) stds = torch.stack([x[1].exp() for x in predictions]) mean = means.mean(0) variance = (stds.square() + means.square()).mean(0) - mean.square() return mean, variance.clamp_min(1e-12).sqrt().log() def to(self, device): for model in self.models: model.to(device) return self def eval(self): for model in self.models: model.eval() return self def parameters(self): for model in self.models: yield from model.parameters() def _concat(batches): return {key: np.concatenate([batch[key] for batch in batches], axis=0) for key in batches[0]} def collect_predictions(model, dataset, device, batch_size=64, stress=None, seed=0): means, log_stds, arrays = [], [], [] with torch.inference_mode(): for i, batch in enumerate(DataLoader(dataset, batch_size=batch_size, shuffle=False)): device_batch = to_device(batch, device) inputs = stress_observations(device_batch, stress, seed + i) if stress else device_batch # Targets remain outside the callable input contract. mean, log_std = model({k: v for k, v in inputs.items() if k not in ("target", "target_mask")}) if stress: # Decode from corrupted observed current, as the deployed API does. anchor_difference = (inputs["current"] - device_batch["current"]).gather( 1, inputs["query_fields"] ) mean = mean + anchor_difference queried_available = inputs["current_mask"].gather(1, inputs["query_fields"]).bool() batch["target_mask"] = batch["target_mask"].bool() & queried_available.cpu() means.append(mean.cpu().numpy()) log_stds.append(log_std.cpu().numpy()) arrays.append({k: v.numpy() for k, v in batch.items()}) if not means: raise ValueError("No benchmark windows") return np.concatenate(means), np.concatenate(log_stds), _concat(arrays) def _families(dataset): return [str(dataset.episodes[ep].family_id) for ep, _ in dataset.indices] def sampling_provenance(dataset, manifest_path): anchors = [(str(dataset.episodes[ep].system_id), int(step)) for ep, step in dataset.indices] families = _families(dataset) return { "manifest_sha256": sha256(manifest_path), "window_sampling": { "seed": dataset.seed, "history_steps": dataset.history_steps, "future_steps": dataset.future_steps, "sampling": "uniform eligible anchors with replacement", "anchor_sha256": hashlib.sha256(json.dumps(anchors, separators=(",", ":")).encode()).hexdigest(), "family_window_counts": {family: families.count(family) for family in sorted(set(families))}, }, } def summarize(mean, log_std, arrays, families, normalization, calibration=None): return metrics_by_group( mean, log_std, arrays["target"], arrays["query_fields"], arrays["query_horizons"], arrays["target_mask"], families, normalization, calibration=calibration, field_names=[spec.name for spec in FIELD_SPECS], ) def profile_inference(model, example, devices=("cpu",), repetitions=30, warmup=5): """Synchronized model-only latency; excludes validation and tensor transfers.""" profiles = [] one = {k: v[:1] for k, v in example.items() if k not in ("target", "target_mask")} for name in devices: if name == "mps" and not torch.backends.mps.is_available(): profiles.append({"device": name, "status": "unavailable"}) continue if hasattr(model, "to"): model.to(name) if hasattr(model, "eval"): model.eval() batch = to_device(one, name) memory_samples, durations = [], [] with torch.inference_mode(): for index in range(warmup + repetitions): if name == "mps": torch.mps.synchronize() start = time.perf_counter_ns() model(batch) if name == "mps": torch.mps.synchronize() memory_samples.append(torch.mps.driver_allocated_memory()) duration = (time.perf_counter_ns() - start) / 1e6 if index >= warmup: durations.append(duration) rss = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss profiles.append( { "device": name, "status": "completed", "batch_size": 1, "precision": "float32", "repetitions": repetitions, "warmup": warmup, "p50_ms": float(np.percentile(durations, 50)), "p95_ms": float(np.percentile(durations, 95)), "process_peak_rss_bytes": int(rss if sys.platform == "darwin" else rss * 1024), "sampled_mps_driver_allocation_peak_bytes": max(memory_samples) if memory_samples else None, "memory_limitation": "RSS is lifetime process high water, includes data and prior models; MPS driver allocation sampled after calls is a lower bound, not resettable model peak.", "timing_scope": "forward only; inputs resident; MPS synchronized before and after every call", } ) return profiles def _candidate_actions(episode, t, steps=8): base = episode.actions[t : t + steps].copy() candidates = [base] if episode.family_id == "thermal": for power in (-5.0, 0.0, 5.0): action = base.copy() action[:, FIELD_INDEX["hvac_power"]] = power candidates.append(action) elif episode.family_id == "tank": for pump in (0.0, 0.5, 1.0): action = base.copy() action[:, FIELD_INDEX["pump_output"]] = pump action[:, FIELD_INDEX["valve_position"]] = 0.7 candidates.append(action) elif episode.family_id == "battery": for power in (-5.0, 0.0, 5.0): action = base.copy() action[:, FIELD_INDEX["battery_power"]] = power candidates.append(action) else: raise ValueError("Counterfactual family has no declared action candidates") return np.asarray(candidates) def evaluate_counterfactual(model, dataset, normalization, device, samples=12): from tfm.simulators import make_simulator records = [] for index in range(min(samples, len(dataset))): epi, t = dataset.indices[index] episode = dataset.episodes[epi] sim = make_simulator(episode.family_id, episode.seed) for j in range(t): sim.step(episode.actions[j], episode.exogenous[j]) snapshot = sim.snapshot() actions = _candidate_actions(episode, t) exogenous = episode.exogenous[t : t + 8] truth_compact = branch_rollouts(sim, snapshot, actions, exogenous) names = [name for name in sim.schema if FIELD_SPECS[FIELD_INDEX[name]].role == "state"] fields = [FIELD_INDEX[name] for name in names] truth = np.zeros((len(actions), 9, len(FIELD_SPECS))) truth[:, :, fields] = truth_compact sample = dataset[index] batch = { key: value.unsqueeze(0).repeat((len(actions),) + (1,) * value.ndim) for key, value in sample.items() if key not in ("target", "target_mask") } normalized_actions = (actions - normalization.mean) / normalization.scale normalized_actions[:, :, ROLE_IDS != 1] = 0 batch["future_actions"] = torch.tensor(normalized_actions, dtype=torch.float32) with torch.inference_mode(): mean, _ = model(to_device(batch, device)) mean = mean.cpu().numpy() horizons = sorted({float(h) for h in sample["query_horizons"].numpy() if h > 0}) observed_current = normalization.inverse(sample["current"].numpy()) for field in fields: if not sample["current_mask"][field]: seen = np.flatnonzero(sample["history_mask"][:, field].numpy()) observed_current[field] = ( ( sample["history"][seen[-1], field].item() * normalization.scale[field] + normalization.mean[field] ) if len(seen) else normalization.mean[field] ) predicted = np.repeat(observed_current[None, None, :], len(actions), axis=0) predicted = np.repeat(predicted, len(horizons), axis=1) lookup = {h: j for j, h in enumerate(horizons)} for q, (field, horizon) in enumerate(zip(sample["query_fields"], sample["query_horizons"])): field, horizon = int(field), float(horizon) if field in fields and horizon in lookup: predicted[:, lookup[horizon], field] += mean[:, q] * normalization.scale[field] selected_steps = np.array([round(h / 15) for h in horizons]) true_at_queries = truth[:, selected_steps] # Sum costs only at declared queried horizons, equally weighted. query_actions = actions[:, selected_steps - 1] predicted_cost = family_cost(episode.family_id, predicted, query_actions, FIELD_INDEX).sum(axis=1) true_cost = family_cost(episode.family_id, true_at_queries, query_actions, FIELD_INDEX).sum(axis=1) result = action_effect_metrics( predicted[:, :, fields], true_at_queries[:, :, fields], predicted_cost, true_cost, field_names=names, horizons=horizons, ) records.append( { "system_id": str(episode.system_id), "family": str(episode.family_id), "anchor_step": int(t), **result, } ) summaries = [] for family in sorted({r["family"] for r in records}): selected = [r for r in records if r["family"] == family] effects = {} for row in selected: for effect in row["effects"]: key = (effect["field"], effect["horizon_minutes"]) effects.setdefault(key, []).append(effect) pairs = sum(row["ranking_comparable_pairs"] for row in selected) summaries.append( { "family": family, "sample_count": len(selected), "mean_regret": float(np.mean([r["regret"] for r in selected])), "ranking_accuracy": sum( (r["ranking_accuracy"] or 0) * r["ranking_comparable_pairs"] for r in selected ) / pairs if pairs else None, "ranking_comparable_pairs": pairs, "effects": [ { "field": field, "horizon_minutes": horizon, "sample_count": len(rows), "action_effect_mae": float(np.mean([r["action_effect_mae"] for r in rows])), "action_effect_rmse": float( np.sqrt(np.mean([r["action_effect_rmse"] ** 2 for r in rows])) ), } for (field, horizon), rows in sorted(effects.items()) ], } ) return { "status": "completed" if records else "unrun", "samples": len(records), "method": "Replay exact simulator to sampled held-out anchor, restore one snapshot for each candidate; identical exogenous trajectory and RNG noise.", "cost_aggregation": "Equal-weight sum at 15,30,60,120 minute queried horizons", "current_reconstruction": "Observed current state; missing fields use last history observation, then training mean. No hidden current state enters predictions.", "oracle_reference": "Simulation-only hindsight cost minimum among supplied candidates using exact shared-noise branches; future noise is available to the oracle, not to learned methods.", "cost_definitions": COST_DEFINITIONS, "candidate_definition": "Observed plan plus constant -5/0/5 kW thermal/battery or 0/0.5/1 pump tank plans; tank valve 0.7.", "records": records, "summary_by_family": summaries, } def _remap_batch(batch, source, destination): result = dict(batch) device = batch["history"].device source_scale = torch.tensor(source.scale, dtype=torch.float32, device=device) source_mean = torch.tensor(source.mean, dtype=torch.float32, device=device) destination_scale = torch.tensor(destination.scale, dtype=torch.float32, device=device) destination_mean = torch.tensor(destination.mean, dtype=torch.float32, device=device) for key, mask_key in (("history", "history_mask"), ("current", "current_mask")): result[key] = ( (batch[key] * source_scale + source_mean - destination_mean) / destination_scale ) * batch[mask_key] for key, role in (("future_actions", 1), ("future_exogenous", 2)): active = batch["field_mask"][:, None, :] * (batch["role_ids"][:, None, :] == role) result[key] = ( (batch[key] * source_scale + source_mean - destination_mean) / destination_scale ) * active return result class SpecialistAdapter: def __init__(self, model, source, destination): self.model, self.source, self.destination = model, source, destination def __call__(self, batch): mean, log_std = self.model(_remap_batch(batch, self.source, self.destination)) ratio = torch.tensor( self.destination.scale / self.source.scale, dtype=mean.dtype, device=mean.device )[batch["query_fields"]] return mean * ratio, log_std + ratio.log() def to(self, device): self.model.to(device) return self def eval(self): self.model.eval() return self def parameters(self): return self.model.parameters() def benchmark(args): manifest, splits = load_manifest_episodes(args.data) if args.counterfactual_samples > 0 and manifest.get("generator_version") != GENERATOR_VERSION: raise ValueError( "Exact counterfactual replay requires data from the installed simulator generator version" ) device = resolve_device(args.device) paths = [Path(p) for p in args.checkpoint] model_members, checkpoints = zip(*(load_checkpoint(path, device) for path in paths)) norm_path = paths[0].parent / "normalization.json" normalization = Normalization.load(norm_path) norm_json = json.loads(norm_path.read_text()) norm_version = norm_json.get("version", "tfm-normalization-v1") for path, checkpoint in zip(paths, checkpoints): if checkpoint.get("normalization_sha256") != sha256(path.parent / "normalization.json"): raise ValueError("Checkpoint normalization fingerprint is missing or does not match its sidecar") if "schema_version" in checkpoint and checkpoint["schema_version"] != SCHEMA_VERSION: raise ValueError("Checkpoint schema/generator version does not match the installed vocabulary") for path in paths[1:]: other = Normalization.load(path.parent / "normalization.json") if not np.array_equal(other.mean, normalization.mean) or not np.array_equal( other.scale, normalization.scale ): raise ValueError("Ensemble members require identical training normalization") model = Ensemble(model_members) if len(model_members) > 1 else model_members[0] reference = checkpoints[0] for checkpoint in checkpoints: for key in ( "training_system_ids", "validation_selection_system_ids", "calibration_system_ids", "test_system_ids", ): if checkpoint.get(key, []) != reference.get(key, []): raise ValueError("Ensemble members need identical recorded partitions") train_ids, test_ids = _ids(splits["train"]), _ids(splits["test"]) if set(test_ids) & ( set(reference.get("training_system_ids", [])) | set(reference.get("validation_selection_system_ids", [])) ): raise ValueError("Benchmark test systems occurred in checkpoint training or model selection") if set(normalization.training_system_ids) & set(test_ids): raise ValueError("Normalization contains test systems") if set(normalization.training_system_ids) != set(reference.get("training_system_ids", [])): raise ValueError("Normalization training systems differ from checkpoint training provenance") history_steps = reference["model_config"].get("history_steps", 24) dataset = WindowDataset( splits["test"], normalization, history_steps=history_steps, future_steps=8, samples=args.samples, seed=args.seed, ) experiment = getattr(args, "experiment", None) if experiment and experiment.axis == "E": selected = [ pair for pair in dataset.indices if dataset.episodes[pair[0]].family_id == experiment.excluded_family ] if not selected: raise ValueError("Leave-family-out test dataset contains no windows of the excluded family") dataset.indices = selected families = _families(dataset) output = Path(args.output) output.mkdir(parents=True, exist_ok=True) calibration_ids = reference.get("calibration_system_ids", []) calibration_episodes = [e for e in splits["validation"] if str(e.system_id) in calibration_ids] calibration_data = ( WindowDataset( calibration_episodes, normalization, history_steps=history_steps, future_steps=8, samples=args.calibration_samples, seed=args.seed + 1, ) if calibration_episodes else None ) reports = {} models = { "foundation_model": model, "persistence": PersistenceBaseline(), "history_linear_system_id": HistoryLinearBaseline(ROLE_IDS), } if args.catboost: training_data = WindowDataset( splits["train"], normalization, history_steps=history_steps, future_steps=8, samples=args.catboost_samples, seed=args.seed + 2, ) training_batch = next(iter(DataLoader(training_data, batch_size=len(training_data)))) numpy_batch = {k: v.numpy() for k, v in training_batch.items()} try: models["catboost"] = CatBoostBaseline(seed=args.seed).fit( numpy_batch, training_system_ids=train_ids, fitted_system_ids=train_ids ) except ImportError: reports["catboost"] = {"status": "unrun", "reason": "Optional catboost package is not installed"} else: reports["catboost"] = { "status": "unrun", "reason": "Enable --catboost to fit the optional training-only baseline", } fingerprints = [sha256(path) for path in paths] for name, candidate in models.items(): calibration = None if calibration_data: cal_mean, cal_log_std, cal_arrays = collect_predictions( candidate, calibration_data, device, args.batch_size ) calibration = fit_calibration( cal_mean, np.exp(cal_log_std), cal_arrays["target"], cal_arrays["query_fields"], cal_arrays["query_horizons"], cal_arrays["target_mask"], calibration_system_ids=_ids(calibration_episodes), training_system_ids=reference.get("training_system_ids", []), validation_selection_system_ids=reference.get("validation_selection_system_ids", []), test_system_ids=test_ids, checkpoint_sha256=fingerprints if name == "foundation_model" else [], normalization_version=norm_version, normalization_sha256=sha256(norm_path), ) save_calibration( calibration, output / ("calibration.json" if name == "foundation_model" else f"calibration_{name}.json"), ) stress = experiment if experiment and experiment.axis == "C" else None mean, log_std, arrays = collect_predictions( candidate, dataset, device, args.batch_size, stress, args.seed ) applied_calibration = None if stress else calibration metric_rows = summarize(mean, log_std, arrays, families, normalization, applied_calibration) calibrated_count = sum( row["count"] for row in metric_rows if row["interval_kind"] == "heldout_calibrated" ) total_count = sum(row["count"] for row in metric_rows) reports[name] = { "status": "completed", "metrics": metric_rows, "calibrated_query_fraction": calibrated_count / total_count if total_count else 0, "calibration_status": "uncalibrated_under_distribution_shift" if stress else "heldout_partial" if calibration and calibrated_count < total_count else "heldout" if calibration else "unavailable_no_disjoint_calibration_systems", } if stress: reports[name]["evaluated_target_fraction"] = float( arrays["target_mask"].sum() / max(1, sum(int(dataset[i]["target_mask"].sum()) for i in range(len(dataset)))) ) reports[name]["missing_current_policy"] = ( "Queries without observed current sensors are excluded, matching API query rejection" ) if not stress: reports[name]["counterfactual"] = evaluate_counterfactual( candidate, dataset, normalization, device, args.counterfactual_samples ) for specification in args.specialist_checkpoint: family, path = specification.split("=", 1) if family not in SCHEMAS or family not in families: reports[f"specialist_{family}"] = { "status": "unrun", "reason": "Family is unknown or absent from sampled benchmark windows", } continue specialist, metadata = load_checkpoint(path, device) if metadata.get("training_families") != [family]: reports[f"specialist_{family}"] = { "status": "unrun", "reason": "Checkpoint provenance does not prove training on this family alone", } continue specialist_norm_path = Path(path).parent / "normalization.json" if metadata.get("normalization_sha256") != sha256(specialist_norm_path): raise ValueError("Specialist normalization fingerprint is missing or does not match") if set(test_ids) & ( set(metadata.get("training_system_ids", [])) | set(metadata.get("validation_selection_system_ids", [])) ): raise ValueError("Specialist checkpoint has test-system leakage") specialist_norm = Normalization.load(specialist_norm_path) adapter = SpecialistAdapter(specialist, normalization, specialist_norm) family_dataset = copy.copy(dataset) family_dataset.indices = [ pair for pair in dataset.indices if dataset.episodes[pair[0]].family_id == family ] mean, log_std, arrays = collect_predictions(adapter, family_dataset, device, args.batch_size) rows = summarize(mean, log_std, arrays, _families(family_dataset), normalization) reports[f"specialist_{family}"] = { "status": "completed", "checkpoint": str(path), "metrics": [r for r in rows if r["family"] == family], "calibration_status": "uncalibrated", "window_alignment": "Identical heldout windows, remapped training normalization", "counterfactual": evaluate_counterfactual( adapter, family_dataset, normalization, device, args.counterfactual_samples ), } if not args.specialist_checkpoint: reports["specialists"] = {"status": "unrun", "reason": "No specialist checkpoint artifacts supplied"} sample_batch = next(iter(DataLoader(dataset, batch_size=1))) profile = profile_inference( model, sample_batch, devices=("cpu", "mps"), repetitions=args.latency_repetitions ) model.to(device) result = { "version": "tfm-benchmark-v1", "seed": args.seed, "samples": len(dataset), "device": str(device), "checkpoint_sha256": fingerprints, "normalization_sha256": sha256(norm_path), "generator_version": manifest.get("generator_version"), "test_system_ids": test_ids, "calibration_system_ids": _ids(calibration_episodes), "parameter_count": sum(p.numel() for p in model.parameters()), "inference": profile, "models": reports, "external_datasets": { "status": "unrun", "reason": "No external physical dataset adapter or checkpoint experiment was supplied; synthetic results only.", }, } manifest_path = Path(args.data) if manifest_path.is_dir(): manifest_path = manifest_path / "manifest.json" result.update(sampling_provenance(dataset, manifest_path)) (output / "benchmark.json").write_text(json.dumps(result, indent=2, allow_nan=False) + "\n") configs = default_experiments() configs[0].data, configs[0].checkpoint = str(args.data), str(paths[0]) if args.stress_tests: for config in configs: if config.axis == "C": config.data, config.checkpoint = str(args.data), str(paths[0]) def evaluate_config(config): if config.axis == "B": return {"benchmark": "benchmark.json", "model": "foundation_model"} mean, log_std, arrays = collect_predictions( model, dataset, device, args.batch_size, config, args.seed ) return { "metrics": summarize(mean, log_std, arrays, families, normalization), "evaluated_target_fraction": float( arrays["target_mask"].sum() / max(1, sum(int(dataset[i]["target_mask"].sum()) for i in range(len(dataset)))) ), "missing_current_policy": "Queries without observed current sensors excluded, matching API query rejection", "interval_status": "uncalibrated_under_distribution_shift", } run_experiments(configs, evaluate_config, output / "generalization.json") return result def parser(): result = argparse.ArgumentParser(description=__doc__) result.add_argument("--data", required=True) result.add_argument("--checkpoint", required=True, action="append") result.add_argument("--output", required=True) result.add_argument("--samples", type=int, default=512) result.add_argument("--calibration-samples", type=int, default=512) result.add_argument("--counterfactual-samples", type=int, default=12) result.add_argument("--latency-repetitions", type=int, default=30) result.add_argument("--batch-size", type=int, default=64) result.add_argument("--seed", type=int, default=2026) result.add_argument("--device", choices=("auto", "cpu", "mps"), default="auto") result.add_argument("--catboost", action="store_true") result.add_argument("--catboost-samples", type=int, default=1024) result.add_argument("--specialist-checkpoint", action="append", default=[], help="family=path") result.add_argument("--stress-tests", action="store_true") return result def main(argv=None): args = parser().parse_args(argv) if min(args.samples, args.calibration_samples, args.batch_size, args.latency_repetitions) < 1: raise ValueError("Sample counts, latency repetitions and batch size must be positive") if args.counterfactual_samples < 0: raise ValueError("Counterfactual sample count must be nonnegative") benchmark(args) if __name__ == "__main__": main()