hf_dataset_example.py

"""Standalone example: consume the Track B HuggingFace dataset.

Self-contained — needs only ``numpy`` + ``datasets`` (no Track B package, no
torch).  Shows how to:
  1. load the dataset and apply the canonical by-video val split,
  2. read one trajectory's three box sources (noisy / kalman / gt),
  3. convert the stored 3x3 rotation to the model's ``rot_6d`` and pack the
     12-d ``box_repr = [center(3), log_dims(3), rot_6d(6)]``,
  4. (optional) build a tiny PyTorch input batch.

    python hf_dataset_example.py --hf_dataset /path/to/track_b_hf_dataset
"""

from __future__ import annotations

import argparse
import json
import os

import numpy as np


# --- rotation convention (MUST match training; see the dataset card) ---------
def matrix_to_rot6d(R: np.ndarray) -> np.ndarray:
    """3x3 rotation (box-local -> camera) -> 6-d = first two ROWS, flattened.
    [...,3,3] -> [...,6]."""
    R = np.asarray(R, np.float32)
    return R[..., :2, :].reshape(*R.shape[:-2], 6)


def rot6d_to_matrix(r6: np.ndarray) -> np.ndarray:
    """Inverse of ``matrix_to_rot6d``: row Gram-Schmidt. [...,6] -> [...,3,3]."""
    r6 = np.asarray(r6, np.float32)
    a1, a2 = r6[..., 0:3], r6[..., 3:6]
    b1 = a1 / np.clip(np.linalg.norm(a1, axis=-1, keepdims=True), 1e-8, None)
    a2 = a2 - np.sum(b1 * a2, axis=-1, keepdims=True) * b1
    b2 = a2 / np.clip(np.linalg.norm(a2, axis=-1, keepdims=True), 1e-8, None)
    b3 = np.cross(b1, b2)
    return np.stack([b1, b2, b3], axis=-2)


def box_repr(center: np.ndarray, dims: np.ndarray, R: np.ndarray) -> np.ndarray:
    """(center[T,3], dims[T,3] FULL extents, R[T,3,3]) -> box_repr[T,12]."""
    log_dims = np.log(np.clip(dims, 1e-4, None))
    return np.concatenate([center, log_dims, matrix_to_rot6d(R)], axis=-1).astype(np.float32)


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--hf_dataset", required=True)
    ap.add_argument("--prior", choices=["noisy", "kalman"], default="noisy",
                    help="which box trajectory to feed the model as the prior input")
    args = ap.parse_args()

    from datasets import load_from_disk

    ds = load_from_disk(args.hf_dataset).with_format("numpy")
    print(f"loaded {len(ds)} trajectories")

    # canonical by-video val split (shared benchmark) — shipped with the dataset
    split_path = os.path.join(args.hf_dataset, "val_video_ids.json")
    val_vids = set()
    if os.path.exists(split_path):
        val_vids = set(json.load(open(split_path))["val_video_ids"])
    is_val = np.array([v in val_vids for v in ds["video_id"]])
    print(f"split by video: train={int((~is_val).sum())}  val={int(is_val.sum())}  "
          f"({len(val_vids)} val videos)")
    # NOTE: split BY VIDEO, never by object — objects from one video must not
    # straddle train/val (they share camera trajectory + scene).

    r = ds[0]
    T = int(r["num_frames"])
    print(f"\nexample trajectory: video={r['video_id']} object={r['object_id']}")
    print(f"  category={r['category']!r}  caption={r['caption']!r}")
    print(f"  frames={T}  keyframes(measured)={int(r['num_keyframes'])}")

    prior_c = r[f"{args.prior}_center"]
    prior_d = r[f"{args.prior}_dims"]
    prior_R = r[f"{args.prior}_R"]
    prior_repr = box_repr(prior_c, prior_d, prior_R)        # [T,12] model input
    gt_repr = box_repr(r["gt_center"], r["gt_dims"], r["gt_R"])  # [T,12] target

    ts_sec = (r["timestamps_ns"].astype(np.float64) - r["timestamps_ns"].min()) / 1e9
    measured = np.asarray(r["measured_mask"], bool)
    print(f"  box_repr (input prior) shape={prior_repr.shape}, gt_repr shape={gt_repr.shape}")
    print(f"  timestamps span {ts_sec.max():.2f}s, measured frames={int(measured.sum())}/{T}")

    # per-frame center error of the prior vs GT (the thing the model reduces)
    err = np.linalg.norm(prior_c - r["gt_center"], axis=-1)
    print(f"  prior center error vs GT: mean={err.mean():.3f}m  max={err.max():.3f}m")

    # verify rot_6d round-trips the stored matrix
    R_rt = rot6d_to_matrix(matrix_to_rot6d(r["gt_R"]))
    print(f"  rot_6d round-trip max err: {np.abs(R_rt - r['gt_R']).max():.2e}")

    # optional: a tiny torch batch (only if torch is available)
    try:
        import torch
        batch = {
            "box_repr": torch.from_numpy(prior_repr)[None],      # [1,T,12]
            "timestamps": torch.from_numpy(ts_sec.astype(np.float32))[None],
            "measured_mask": torch.from_numpy(measured)[None],
        }
        print(f"\n  torch batch ready: box_repr {tuple(batch['box_repr'].shape)} "
              f"-> feed to a bidirectional transformer (see trackb/model.py)")
    except ImportError:
        print("\n  (torch not installed — skipping the tensor-batch demo)")


if __name__ == "__main__":
    main()