scripts/smoke_train.py

from __future__ import annotations

from pathlib import Path

import numpy as np
import pandas as pd
import yaml

from fall.train import train_one


def main() -> None:
    root = Path("outputs/smoke")
    pose_dir = root / "poses"
    pose_dir.mkdir(parents=True, exist_ok=True)
    rows = []
    rng = np.random.default_rng(42)
    for i in range(24):
        label = int(i % 2 == 0)
        pose = rng.normal(0, 1, size=(32, 17, 3)).astype(np.float32)
        pose[..., 2] = 0.95
        if label:
            pose[:, :, 1] += np.linspace(0, 3, 32)[:, None]
        path = pose_dir / f"sample_{i:03d}.npy"
        np.save(path, pose)
        split = "train" if i < 16 else "val" if i < 20 else "test"
        rows.append({"video_id": f"sample_{i:03d}", "video_path": "synthetic", "pose_path": str(path), "label": label, "dataset": "synthetic", "split": split})
    manifest = root / "manifest.csv"
    pd.DataFrame(rows).to_csv(manifest, index=False)
    cfg = {
        "seed": 42,
        "device": "cuda",
        "output_dir": str(root / "run"),
        "data": {"type": "pose", "manifest": str(manifest), "frames": 32, "use_confidence": True, "use_velocity": True},
        "model": {"name": "pose_tcn_attention", "hidden_dim": 32, "layers": 2, "dropout": 0.1},
        "train": {"epochs": 2, "batch_size": 8, "eval_batch_size": 8, "num_workers": 0, "lr": 0.001, "class_weight": True, "amp": True, "patience": 2},
    }
    cfg_path = root / "config.yaml"
    with cfg_path.open("w", encoding="utf-8") as f:
        yaml.safe_dump(cfg, f)
    print(train_one(cfg_path))


if __name__ == "__main__":
    main()