SparseWake/sparsewake
SparseWake SparseWake is a synthetic benchmark for sparse temporal hydrodynamic sensing. ICLR 2027 release The expanded release adds controlled multi-source mixtures and common-prior nearest-source tasks, with complete core data banks, reference checkpoints, a small review supplement, and reproduction code with a frozen wake-library input. Download release iclr2027-v1.0rc2 The version page lists the three archives, exact sizes, checksums, extraction instructions… See the full description on the dataset page: https://huggingface.co/datasets/SparseWake/sparsewake.
0275
1from __future__ import annotations2 3import argparse4import json5from pathlib import Path6import sys7 8import numpy as np9import yaml10 11ROOT = Path(__file__).resolve().parents[1]12sys.path.insert(0, str(ROOT / "src"))13 14from sparsewake.data import load_h515from sparsewake.evaluate import evaluate_predictions, predict16from sparsewake.features import build_design_matrix17from sparsewake.splits import pose_holdout_split18from sparsewake.train import standardize_train_val_test, train_temporal_mlp19 20 21def main() -> None:22 parser = argparse.ArgumentParser()23 parser.add_argument("--config", required=True)24 parser.add_argument("--data", default=None)25 parser.add_argument("--quick", action="store_true")26 parser.add_argument("--out", default="tables/quick_train_metrics.json")27 args = parser.parse_args()28 cfg = yaml.safe_load(Path(args.config).read_text())29 data_path = Path(args.data) if args.data else ROOT / cfg["data"]30 data = load_h5(data_path, input_key=cfg.get("input_key", "X_raw"))31 history = 4 if args.quick else int(cfg.get("history", 24))32 x, idx = build_design_matrix(data, feature_set=cfg.get("feature_set", "raw_norm"), history=history)33 y = data["target"][idx]34 pose_id = data["pose_id"][idx]35 train_idx, val_idx, test_idx = pose_holdout_split(pose_id, seed=int(cfg.get("seed", 1)))36 if args.quick:37 train_idx = train_idx[: min(len(train_idx), 512)]38 val_idx = val_idx[: min(len(val_idx), 128)]39 test_idx = test_idx[: min(len(test_idx), 128)]40 x, _, _ = standardize_train_val_test(x, train_idx)41 output_dim = 3 if cfg.get("target", "location") == "location_theta" else 242 model = train_temporal_mlp(43 x,44 y,45 train_idx,46 val_idx,47 output_dim=output_dim,48 epochs=3 if args.quick else int(cfg.get("epochs", 50)),49 batch_size=int(cfg.get("batch_size", 1024)),50 seed=int(cfg.get("seed", 1)),51 )52 pred = predict(model, x[test_idx])53 metrics = evaluate_predictions(y[test_idx, :output_dim], pred)54 metrics["quick_mode"] = bool(args.quick)55 metrics["n_train"] = int(len(train_idx))56 metrics["n_test"] = int(len(test_idx))57 out = ROOT / args.out58 out.parent.mkdir(parents=True, exist_ok=True)59 out.write_text(json.dumps(metrics, indent=2) + "\n")60 print(json.dumps(metrics, indent=2))61 62 63if __name__ == "__main__":64 main()65 66