| """Reload the saved model in a fresh process and predict all 922 test cases.""" |
| import argparse |
| from pathlib import Path |
| import time |
|
|
| import numpy as np |
| import pandas as pd |
| from common import (load_config, ensure_compute_node, load_fixed_dataset, WavyChannelRF, |
| metrics_for, print_metrics, prediction_frame, write_json, write_metrics, sha256) |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser(description=__doc__) |
| parser.add_argument("--config", required=True) |
| parser.add_argument("--checkpoint", help="Optional explicit path to saved joblib .pth bundle") |
| args = parser.parse_args() |
| config = load_config(args.config) |
| runtime = ensure_compute_node(config) |
| directory = Path(config["output"]["results_dir"]) |
| data, data_manifest = load_fixed_dataset(config) |
| test = data.loc[data[config["data"]["split_column"]] == "test"] |
| features, targets = config["data"]["features"], config["data"]["targets"] |
| checkpoint = Path(args.checkpoint or config["output"]["checkpoint"]) |
| print(f"[inference] Loading {checkpoint}", flush=True) |
| model, metadata = WavyChannelRF.load(checkpoint) |
| if model.n_train_ != config["data"]["expected_counts"]["train"]: |
| raise ValueError("Checkpoint did not fit the required complete training set") |
| for name in ["dataset", "split_file"]: |
| if metadata["data"]["files"][name]["sha256"] != data_manifest["files"][name]["sha256"]: |
| raise ValueError(f"{name} changed since model fitting") |
| start = time.perf_counter() |
| predicted = model.predict(test[features]) |
| seconds = time.perf_counter() - start |
| metrics = metrics_for(test[targets], predicted, targets) |
| print_metrics("inference-test", metrics, len(test)) |
| output = prediction_frame(test, predicted, config) |
| output.to_csv(directory / "inference_predictions.csv", index=False) |
| write_metrics(directory, metrics, "inference_metrics") |
| reference = pd.read_csv(directory / "test_predictions.csv") |
| key = config["data"]["join_key"] |
| if reference[key].tolist() != output[key].tolist(): |
| raise ValueError("Reloaded inference test case order differs from training evaluation") |
| pred_cols = [f"{t}_pred" for t in targets] |
| error = float(np.max(np.abs(reference[pred_cols].to_numpy() - output[pred_cols].to_numpy()))) |
| np.testing.assert_allclose(reference[pred_cols], output[pred_cols], atol=1e-10, rtol=1e-12) |
| verification = {"status": "passed", "test_rows": len(test), "predictions_finite": True, |
| "independent_process_reload": True, "case_uids_equal": True, |
| "max_abs_prediction_difference": error, "tolerance": {"atol": 1e-10, "rtol": 1e-12}, |
| "prediction_seconds": seconds, "runtime": runtime, |
| "checkpoint_sha256": sha256(checkpoint), "test_metrics": metrics} |
| write_json(directory / "reload_validation.json", verification) |
| print(f"[inference] Reload validation passed: n={len(test)}, max_abs_difference={error:.3g}", flush=True) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
