load_failbench.py

"""Standalone loader for the FailBench v2 contact-prediction datasets.

Pure ``h5py + numpy + scipy`` (+ optional ``torch`` for the Dataset wrapper) — it has
ZERO dependency on the FailBench package, MuJoCo, or robosuite. Use it to load a trial and
rebuild its agentview contact-heatmap target straight from the published HDF5 files.

    pip install h5py hdf5plugin numpy scipy pandas        # torch optional, for the Dataset
    python load_failbench.py --data_root /path/to/failbench-robocasa-v2 \
                             --cache_root /path/to/failbench-robocasa-v2/target_cache

============================  THE ONE GOTCHA  ============================
The arrays are Blosc(lz4)-compressed with the third-party HDF5 filter id 32001.
You MUST ``import hdf5plugin`` BEFORE h5py opens any file, or reads fail with
"can't open plugin directory". (``dataset.compression`` misleadingly reports
``None`` for this filter — the data IS compressed.) This module does it at the
top, so always import this module (or hdf5plugin) before opening the files.
=========================================================================

Repo layout this loader understands (either dataset):

    <data_root>/
      v2/<task>.h5                       # RoboCasa: task files directly under v2/
      v2/manifest_train.csv              # RoboCasa train split (+ manifest.csv, quarantine.csv)
      v2/<split>/<task>.h5               # LIBERO: split = libero_spatial|object|goal
      v2/<split>/manifest.csv
    <cache_root>/                        # == <data_root>/target_cache
      <split>/<task>.h5                  # per-trial /<trial_id>/projection (N,3)[u,v,force] + failure_prob attr
"""
from __future__ import annotations

import argparse
from pathlib import Path

import hdf5plugin  # noqa: F401  — registers the Blosc filter; MUST precede h5py file open
import h5py
import numpy as np
from scipy.ndimage import gaussian_filter

# Default agentview grid (H, W). cam_agentview_size is stored as [W, H] = [320, 240].
DEFAULT_HW = (240, 320)

# The per-trial dataset keys written by the pipeline. read_trial() returns whatever subset
# of these (plus scalar/array attrs) a given trial actually has.
_DATASET_KEYS = (
    # 8-frame pre-failure window (T=8, 4 fps), agentview + wrist RGB/depth + proprio
    "window_frame_idx", "window_qpos", "window_qvel", "window_ee_pos", "window_gripper_ctrl",
    "window_agentview_rgb", "window_agentview_depth", "window_wrist_rgb", "window_wrist_depth",
    # goal-conditioning frames
    "goal_qpos", "goal_qvel", "goal_ee_pos", "goal_gripper_ctrl", "goal_offsets",
    # single pre-failure frame (v1-compatible)
    "pre_qpos", "pre_qvel", "pre_ee_pos", "pre_gripper_ctrl", "pre_target_qpos",
    "pre_rgb", "pre_depth", "robot0_eye_in_hand_rgb", "robot0_eye_in_hand_depth",
    # contacts recorded over the post-failure settle (world frame)
    "contact_positions", "contact_forces", "contact_force_world", "contact_time",
    "contact_geom_pairs", "contact_failure_id", "impacted_geom_ids", "geom_bodyid",
    # baseline (healthy-hold replay) contacts — used by the failure-induced filter
    "baseline_contact_geom_pairs", "baseline_contact_positions",
    # post-failure observations
    "post_agentview_rgb", "post_agentview_depth", "post_wrist_rgb", "post_wrist_depth",
    # agentview camera calibration (lets you project contacts yourself, no sim needed)
    "cam_agentview_pos", "cam_agentview_mat0", "cam_agentview_fovy", "cam_agentview_size",
    "cam_wrist_pos_window", "cam_wrist_mat0_window", "cam_wrist_fovy", "cam_wrist_size",
    # failure descriptor + object poses + settle state trajectory
    "failure_joints", "obj_names", "obj_pos_pre", "obj_quat_pre", "obj_pos_post", "obj_quat_post",
    "settle_step_idx", "settle_qpos", "settle_qvel", "settle_gripper_qpos",
    "settle_obj_pos", "settle_obj_quat",
)
_SCALAR_ATTRS = (
    "trial_id", "split", "task", "demo_key", "seed", "seed_idx", "bin_idx", "fail_idx",
    "traj_progress", "failure_mode", "failure_prob", "is_holding", "force_frame", "scene_table_z",
)
_ARRAY_ATTRS = ("scene_aabb_min", "scene_aabb_max", "robot_geom_ids", "scene_entities_json")


# --------------------------------------------------------------------------- paths
def resolve_h5(data_root, split: str, task: str) -> Path:
    """Locate a task's v2 HDF5 by (split, task). Handles both repo layouts.

    NOTE: do NOT trust the manifest's ``h5_path`` column — it stores the absolute path
    of the original build machine. Always resolve relative to your local repo root.
    """
    root = Path(data_root) / "v2"
    for cand in (root / f"{task}.h5", root / split / f"{task}.h5"):
        if cand.exists():
            return cand
    raise FileNotFoundError(f"no v2 h5 for split={split} task={task} under {root}")


def cache_h5(cache_root, split: str, task: str) -> Path:
    return Path(cache_root) / split / f"{task}.h5"


def load_manifest(data_root):
    """Concatenate every manifest under ``v2/`` into one DataFrame (requires pandas).

    Prefers ``manifest_train.csv`` (RoboCasa train split); falls back to per-split
    ``manifest.csv`` (LIBERO). Adds no rows for ``quarantine.csv``.
    """
    import pandas as pd
    root = Path(data_root) / "v2"
    paths = sorted(root.rglob("manifest_train.csv")) or sorted(root.rglob("manifest.csv"))
    if not paths:
        raise FileNotFoundError(f"no manifest*.csv under {root}")
    return pd.concat([pd.read_csv(p) for p in paths], ignore_index=True)


# --------------------------------------------------------------------------- read
def read_trial(h5_path, trial_id: str, keys=None) -> dict:
    """Load one trial group into a dict of numpy arrays + scalar/array attrs."""
    want = set(keys) if keys is not None else set(_DATASET_KEYS)
    out: dict = {}
    with h5py.File(h5_path, "r", libver="latest", swmr=True) as f:
        grp = f[f"trials/{trial_id}"]
        for k in want:
            if k in grp:
                ds = grp[k]
                if h5py.check_string_dtype(ds.dtype) is not None:
                    out[k] = [s.decode() if isinstance(s, bytes) else s for s in ds[:]]
                else:
                    out[k] = ds[()]
        for k in _SCALAR_ATTRS + _ARRAY_ATTRS:
            if k in grp.attrs:
                v = grp.attrs[k]
                out[k] = v.decode() if isinstance(v, bytes) else v
    return out


def trial_ids(h5_path) -> list:
    with h5py.File(h5_path, "r", libver="latest", swmr=True) as f:
        return sorted(f["trials"].keys()) if "trials" in f else []


# --------------------------------------------------------------------------- targets
def heatmap_from_projection(cache_root, split: str, task: str, trial_id: str, *,
                            hw=DEFAULT_HW, sigma_px: float = 4.0, log1p: bool = False) -> np.ndarray:
    """Rebuild the (H, W) contact-mass heatmap from the precomputed projection cache.

    The cache stores, per trial, an (N, 3) array of in-frame ``[u, v, force_mag]`` plus a
    scalar ``failure_prob`` attr (the contacts are already failure-induced-filtered). The
    dense target is ``scatter(force * failure_prob) -> Gaussian blur``. This matches the
    GPU training target (``build_target_from_projection``); pass ``log1p=True`` to match
    that path's default compression of the mass.
    """
    H, W = hw
    out = np.zeros((H, W), np.float32)
    with h5py.File(cache_h5(cache_root, split, task), "r") as f:
        g = f[trial_id]
        proj = np.asarray(g["projection"], np.float32)        # (N, 3) [u, v, force]
        fp = float(g.attrs.get("failure_prob", 1.0))
    if proj.shape[0]:
        u = np.clip(proj[:, 0].astype(int), 0, W - 1)
        v = np.clip(proj[:, 1].astype(int), 0, H - 1)
        np.add.at(out, (v, u), proj[:, 2] * fp)
        if sigma_px > 0:
            out = gaussian_filter(out, sigma=sigma_px, mode="constant", cval=0.0)
    return np.log1p(out) if log1p else out


def heatmap_from_contacts(trial: dict, *, sigma_px: float = 4.0,
                          weighting: str = "force_prior") -> np.ndarray:
    """Rebuild the heatmap on-the-fly from raw contacts + camera calibration (no cache, no sim).

    Mirrors the reference projection exactly: the agentview camera looks down its own −Z; the
    image v-axis points down, so the rotation's Y-row is negated. ``weighting`` is one of
    ``force_prior`` (force × failure_prob), ``force``, or ``count``. Note this uses the raw
    recorded contacts (no failure-induced filtering); use the projection cache for the filtered
    training target.
    """
    cam_pos = np.asarray(trial["cam_agentview_pos"], np.float64)
    R = np.asarray(trial["cam_agentview_mat0"], np.float64).reshape(3, 3).T.copy()
    R[1] = -R[1]
    W, H = (int(x) for x in trial["cam_agentview_size"])
    fy = (H / 2.0) / np.tan(np.deg2rad(float(trial["cam_agentview_fovy"])) / 2.0)
    fx, cx, cy = fy, W / 2.0, H / 2.0

    out = np.zeros((H, W), np.float32)
    pts = np.asarray(trial.get("contact_positions", np.zeros((0, 3))), np.float64)
    if pts.shape[0] == 0:
        return out
    p_cam = (pts - cam_pos) @ R.T
    depth = -p_cam[:, 2]
    front = depth > 1e-6
    if not front.any():
        return out
    u = fx * p_cam[front, 0] / depth[front] + cx
    v = fy * p_cam[front, 1] / depth[front] + cy
    inb = (u >= 0) & (u < W) & (v >= 0) & (v < H)
    if not inb.any():
        return out

    if weighting == "count":
        w_all = np.ones(pts.shape[0], np.float32)
    else:
        f = trial.get("contact_force_world")
        if f is None:
            f = np.asarray(trial["contact_forces"])[:, :3]
        w_all = np.linalg.norm(np.asarray(f, np.float32), axis=1)
        if weighting == "force_prior":
            w_all = w_all * np.float32(trial.get("failure_prob", 1.0))
    w = w_all[front][inb]
    np.add.at(out, (np.clip(v[inb].astype(int), 0, H - 1), np.clip(u[inb].astype(int), 0, W - 1)), w)
    return gaussian_filter(out, sigma=sigma_px, mode="constant", cval=0.0) if sigma_px > 0 else out


# --------------------------------------------------------------------------- torch Dataset
def make_dataset(data_root, cache_root=None, *, hw=DEFAULT_HW, sigma_px=4.0, log1p=False):
    """Return a ``torch.utils.data.Dataset`` over the manifest.

    Each item is ``{"trial": <dict>, "target": (H,W) float32}``. When ``cache_root`` is given
    the target comes from the projection cache (filtered, matches training); otherwise it is
    rebuilt on-the-fly from the raw contacts.
    """
    import torch

    df = load_manifest(data_root)

    class _FailBenchV2(torch.utils.data.Dataset):
        def __len__(self):
            return len(df)

        def __getitem__(self, i):
            row = df.iloc[i]
            split, task, tid = row["split"], row["task"], row["trial_id"]
            trial = read_trial(resolve_h5(data_root, split, task), tid)
            if cache_root is not None:
                tgt = heatmap_from_projection(cache_root, split, task, tid,
                                              hw=hw, sigma_px=sigma_px, log1p=log1p)
            else:
                tgt = heatmap_from_contacts(trial, sigma_px=sigma_px)
            return {"trial": trial, "target": torch.from_numpy(np.ascontiguousarray(tgt))}

    return _FailBenchV2()


# --------------------------------------------------------------------------- demo / self-test
def _first_filtered_trial(cache_root, split, task):
    """First trial id in this task's cache whose (filtered) projection is non-empty."""
    p = cache_h5(cache_root, split, task)
    if not p.exists():
        return None
    with h5py.File(p, "r") as f:
        for tid in f.keys():
            if f[tid]["projection"].shape[0] > 0:
                return tid
    return None


def _demo(data_root, cache_root):
    df = load_manifest(data_root)
    print(f"manifest rows: {len(df)} | tasks: {df['task'].nunique()} | splits: {sorted(df['split'].unique())}")
    # pick a trial whose FAILURE-INDUCED (filtered) target is non-empty, so both the
    # on-the-fly and the cached heatmaps carry mass and the demo is meaningful.
    contact_rows = df[df["n_contacts"] > 0] if "n_contacts" in df.columns else df
    row = contact_rows.iloc[0]
    if cache_root:
        for _, r in contact_rows.iterrows():
            tid_f = _first_filtered_trial(cache_root, r["split"], r["task"])
            if tid_f is not None:
                row = r.copy(); row["trial_id"] = tid_f
                break
    split, task, tid = row["split"], row["task"], row["trial_id"]
    h5 = resolve_h5(data_root, split, task)
    print(f"\nloading {split}/{task}/{tid}  from {h5.name}")
    trial = read_trial(h5, tid)
    for k in ("window_agentview_rgb", "pre_rgb", "contact_positions", "settle_qpos"):
        if k in trial:
            print(f"  {k:24s} {np.asarray(trial[k]).shape} {np.asarray(trial[k]).dtype}")
    print(f"  failure_mode={trial.get('failure_mode')} failure_prob={trial.get('failure_prob')}")

    live = heatmap_from_contacts(trial)
    print(f"\non-the-fly heatmap: shape={live.shape} sum={live.sum():.3f} max={live.max():.4f}")
    assert live.sum() > 0, "expected non-zero contact mass on a contact-bearing trial"
    if cache_root and cache_h5(cache_root, split, task).exists():
        cached = heatmap_from_projection(cache_root, split, task, tid)
        print(f"cached (filtered) heatmap: shape={cached.shape} sum={cached.sum():.3f} max={cached.max():.4f}")
    print("\nOK — load + target rebuild verified.")


if __name__ == "__main__":
    ap = argparse.ArgumentParser(description=__doc__)
    ap.add_argument("--data_root", required=True, help="repo root containing v2/")
    ap.add_argument("--cache_root", default=None, help="target_cache/ dir (optional)")
    args = ap.parse_args()
    _demo(args.data_root, args.cache_root)