"""Reload saved weights and evaluate 200 original-scale 32x32 pressure fields.""" import csv from copy import deepcopy from pathlib import Path import torch from common import (arguments, atomic_json, build_model, configure_runtime, evaluate, load_config, load_data, relative_l2, sha256) def main(): parser = arguments(__doc__) parser.add_argument("--checkpoint", type=Path) args = parser.parse_args() config = load_config(args.config) device = configure_runtime(config) output = Path(config["output"]["root"]) results = output / config["output"]["results"] checkpoint_path = args.checkpoint or output / config["output"]["checkpoint"] last = torch.load(output / config["output"]["last_checkpoint"], map_location="cpu", weights_only=False) assert last["epoch"] == config["training"]["epochs"] == 500 checkpoint = torch.load(checkpoint_path, map_location="cpu", weights_only=False) trained_config = deepcopy(checkpoint["config"]) trained_config["data"]["root"] = config["data"]["root"] trained_config["output"]["root"] = config["output"]["root"] if trained_config != config: raise ValueError("Inference configuration does not match trained checkpoint") data, manifest = load_data(config) for split, info in manifest["files"].items(): assert info["sha256"] == checkpoint["data_sha256"][split], "Dataset changed after training" model = build_model(config, device) model.load_state_dict(checkpoint["model_state_dict"], strict=True) metrics, errors, predictions = evaluate(model, data["test_x"], data["test_y"], checkpoint["normalizer"], config["training"]["batch_size"], device, return_predictions=True) # Recompute using saved physical fields in float64, avoiding batch-average or # normalized-target metric ambiguities in the final reported value. errors64 = relative_l2(predictions.double(), data["test_y"].double()) metrics["relative_l2_mean"] = errors64.mean().item() metrics["squared_relative_l2_mean"] = errors64.square().mean().item() metrics["relative_l2_std"] = errors64.std(correction=0).item() reference = config["paper_reference"] metrics.update({"n_test": len(errors), "resolution": [32, 32], "scale": "original pressure", "aggregation": "arithmetic mean of per-sample Euclidean norm ratios", "completed_epochs": last["epoch"], "selected_epoch": checkpoint["epoch"], "checkpoint": str(checkpoint_path), "checkpoint_sha256": sha256(checkpoint_path), "checkpoint_selection": checkpoint["selection_rule"], "paper_table2": reference, "signed_gap": metrics["relative_l2_mean"] - reference, "absolute_gap": abs(metrics["relative_l2_mean"] - reference), "relative_gap_percent": 100 * (metrics["relative_l2_mean"] / reference - 1), "last_epoch_test_relative_l2": last["test_metrics"]["relative_l2_mean"], "test_used_for_selection": False, "loss_metric_caveat": "Paper Eq15 is squared; this report's primary metric is unsquared as requested"}) torch.save({"x": data["test_x"], "y": data["test_y"], "prediction": predictions, "sample_ids": torch.arange(len(errors)), "relative_l2": errors64, "checkpoint_sha256": metrics["checkpoint_sha256"]}, results / "predictions.pt") with (results / "per_sample_metrics.csv").open("w", newline="") as stream: writer = csv.writer(stream) writer.writerow(["sample_id", "relative_l2_original_scale", "squared_relative_l2_original_scale"]) for index, error in enumerate(errors64.tolist()): writer.writerow([index, error, error * error]) atomic_json(results / "metrics.json", metrics) print(f"RELOADED {checkpoint_path} epoch={checkpoint['epoch']}; actual training=500 epochs", flush=True) print(f"INFERENCE original-scale per-sample relative L2 mean={metrics['relative_l2_mean']:.10f}; " f"paper=0.0351; signed_gap={metrics['signed_gap']:+.10f}; " f"relative_gap={metrics['relative_gap_percent']:+.3f}%", flush=True) print(f"Squared relative L2 mean (paper Eq15)={metrics['squared_relative_l2_mean']:.10f}", flush=True) if __name__ == "__main__": main()