React / preprocess /backfill.py
yxma's picture
Add tactile_{left,right}_is_new flags; ship preprocess/ code; document true tactile rate (71.8% duplicated, ~8.5 fps effective)
e37c13d verified
Raw
History Blame Contribute Delete
7.7 kB
"""Add tactile validity flags to parquet files that were built before them.
The already-published release has no ``tactile_*_is_new`` columns, and
rebuilding it from source would mean re-encoding every video. Instead we
recover the flags from data already in the parquet.
Method — a repeated GelSight frame produces a bit-identical contact triple
(intensity, area, mixed), because all three are deterministic reductions of
the same pixels. So a row is a fresh reading exactly when its triple differs
from the previous row's.
This is a *proxy*, not a pixel comparison: two genuinely different frames
would have to agree in all three float32 reductions to be missed, which does
not happen in practice but is not impossible. ``verify_against_video()``
checks the proxy against real decoded frames on a sample.
"""
from __future__ import annotations
from pathlib import Path
import numpy as np
import pyarrow.parquet as pq
from .contact import duplication_stats
from .meta import backfill_is_new
SCALARS = ("intensity", "area", "mixed")
def flags_from_scalars(table, side: str) -> np.ndarray:
"""Recover per-row 'this is a new tactile frame' flags for one side."""
cols = [np.asarray(table[f"tactile_{side}_{s}"].to_numpy(), np.float64)
for s in SCALARS if f"tactile_{side}_{s}" in table.column_names]
if not cols:
raise KeyError(f"parquet has no tactile_{side}_* contact columns")
stacked = np.stack(cols, axis=1)
is_new = np.ones(len(stacked), dtype=bool)
if len(stacked) > 1:
is_new[1:] = np.any(stacked[1:] != stacked[:-1], axis=1)
return is_new
def process_parquet(path: Path, dry_run: bool = False) -> dict:
"""Backfill one parquet in place; returns its duplication stats."""
table = pq.read_table(str(path))
left = flags_from_scalars(table, "left")
right = flags_from_scalars(table, "right")
if not dry_run:
pq.write_table(backfill_is_new(table, left, right), str(path))
return {
"path": str(path),
"left": duplication_stats(left),
"right": duplication_stats(right),
}
def process_tree(root: Path, dry_run: bool = False) -> list[dict]:
"""Backfill every ``meta/**/episode_*.parquet`` under a task directory."""
files = sorted(Path(root).rglob("meta/**/episode_*.parquet"))
if not files:
files = sorted(Path(root).rglob("episode_*.parquet"))
return [process_parquet(p, dry_run) for p in files]
def aggregate(reports: list[dict]) -> dict:
"""Dataset-level duplication summary across episodes."""
total = unique = 0
worst = 0
for rep in reports:
for side in ("left", "right"):
s = rep[side]
total += s["n_frames"]
unique += s["n_unique"]
worst = max(worst, s["max_repeat_run"])
ratio = 1.0 - unique / total if total else 0.0
return {
"episodes": len(reports),
"rows": total,
"unique_tactile": unique,
"duplicate_ratio": ratio,
"effective_fps": 30.0 * (1.0 - ratio),
"max_repeat_run": worst,
}
def _source_is_new(h5_path: Path, side: str, start: int, count: int) -> np.ndarray:
"""Bit-exact 'this frame differs from the previous one' over source pixels."""
import h5py
import hdf5plugin # noqa: F401
with h5py.File(str(h5_path), "r") as f:
block = f[f"gelsight/{side}/frames"][start:start + count]
truth = np.ones(len(block), bool)
for i in range(1, len(block)):
truth[i] = not np.array_equal(block[i], block[i - 1])
return truth
def verify_against_h5(parquet_path: Path, h5_path: Path, side: str = "left",
limit: int = 600, shift: int | None = None,
search: range | None = None) -> dict:
"""Ground-truth check of the flags against the source H5 pixels.
The source frames are the only bit-exact reference — the published MP4s are
H.264-encoded, so a duplicated frame does not decode back to identical
pixels (use ``verify_against_video``, which compares with a tolerance).
A published episode may have had a constant tactile latency shift baked in,
in which case parquet row ``i`` corresponds to source frame
``trim + i + shift``. Pass ``shift``, or leave it None to search for the
value that lines the two up — that search doubles as an integrity check
that the intended correction really was applied.
"""
table = pq.read_table(str(parquet_path))
proxy = flags_from_scalars(table, side)[:limit]
trim = int(np.asarray(table["source_h5_frame"].to_numpy())[0])
n_req = len(proxy)
candidates = ([shift] if shift is not None
else list(search) if search is not None else [0, 15])
lo, hi = min(candidates), max(candidates)
# Read the source span once and slide over it — re-reading per candidate
# shift would multiply the (large) HDF5 traffic by len(candidates).
start = max(0, trim + lo - 1) # one extra for a predecessor
pad = (trim + lo) - start # 1 unless clamped at 0
span = _source_is_new(h5_path, side, start, (hi - lo) + n_req + pad)
def slice_truth(sh: int) -> np.ndarray:
off = pad + (sh - lo)
return span[off:off + n_req]
# Row 0 is True by convention on both sides (neither has a predecessor
# inside its own window), so it carries no evidence — compare from row 1.
scored = []
for sh in candidates:
truth = slice_truth(sh)
n = min(len(truth), n_req)
if n < 2:
continue
scored.append((int((truth[1:n] != proxy[1:n]).sum()), n, sh))
if not scored:
raise ValueError(f"{parquet_path.name}: no overlap with source frames")
mismatches, n, best = min(scored, key=lambda x: x[0])
truth = slice_truth(best)
return {
"compared": int(n - 1),
"mismatches": int(mismatches),
"shift": int(best),
"shift_detected": shift is None,
"proxy_unique": int(proxy[1:n].sum()),
"source_unique": int(truth[1:n].sum()),
}
def verify_against_video(parquet_path: Path, video_path: Path,
limit: int = 300, tol: float = 1.0) -> dict:
"""Check the proxy against decoded video, comparing with a tolerance.
H.264 is lossy, so a duplicated source frame still decodes to slightly
different pixels. We therefore call two decoded frames "the same" when
their mean absolute difference stays below `tol`, and report the observed
separation so the threshold can be sanity-checked rather than trusted.
"""
import av
table = pq.read_table(str(parquet_path))
side = "left" if "tactile_left" in video_path.name else "right"
proxy = flags_from_scalars(table, side)[:limit]
mad, prev = [], None
with av.open(str(video_path)) as container:
for i, frame in enumerate(container.decode(video=0)):
if i >= limit:
break
arr = frame.to_ndarray(format="rgb24").astype(np.float32)
mad.append(0.0 if prev is None
else float(np.abs(arr - prev).mean()))
prev = arr
mad = np.asarray(mad)
truth = mad > tol
truth[0] = True
n = min(len(truth), len(proxy))
dup_mad = mad[1:n][~proxy[1:n]]
new_mad = mad[1:n][proxy[1:n]]
return {
"compared": int(n),
"mismatches": int((truth[:n] != proxy[:n]).sum()),
"proxy_unique": int(proxy[:n].sum()),
"video_unique": int(truth[:n].sum()),
"mad_duplicate_max": float(dup_mad.max()) if len(dup_mad) else 0.0,
"mad_new_min": float(new_mad.min()) if len(new_mad) else 0.0,
}