PULSE-code / pulse /tasks /t2_data.py
velvet-pine-22's picture
PULSE-code: benchmark code for the PULSE dataset (camera-ready release)
dffbc1f verified
Raw History Blame Contribute Delete
16.4 kB
"""T2 samples: sliding windows (primary) and annotated segments (secondary).
Timeline
Annotations are in seconds on the recording's alignment timeline
(``n_samples`` frames at 100 Hz). Every benchmark stream of a recording is
placed on that timeline: a stream with ``n_samples`` (or ``n_samples + 1``)
rows starts at frame 0; in v1.1 the EMG, eye-tracking and pressure files of
v2s6, v13s2 and v13s3 were re-trimmed for the release to the MoCap capture
window ``[front, front + n)`` (``alignment_metadata.json``,
``mocap_padding``) and start at frame ``front``. Any other row count raises.
Evaluation span
Per recording, the frames covered by *every* benchmark modality
(MoCap, EMG, eye tracking, IMU, pressure) that the availability table marks
usable, with the start rounded up to the 20 Hz grid. It is computed from
the frozen availability table and the alignment metadata only, so it is the
same for every model row whatever its input modalities: two rows that load
a recording are scored on identical windows / segments, and the common
test set of the aggregator compares like with like.
Signals
Each modality is loaded with the strict loaders (``pulse.data.io``),
checked against the availability row count, cut to the span and decimated
to 20 Hz (every 5th sample; the grid is aligned to the alignment timeline).
Z-scoring uses per-modality mean / SD over the frames of the training
samples (each frame counted once per sample that contains it), applied to
train, inner-validation and test alike.
Labels (frozen taxonomy v1.1)
A segment is kept if ``Taxonomy.classify`` keeps it and it lasts at least
``segment.min_sec``. Sliding protocol: windows of ``window_sec`` at
``stride_sec`` over the span; a window takes the labels of the kept segment
that contains its centre frame and is skipped if there is none. Segment
protocol: one sample per kept segment inside the span, its own
``[start, end)``; segments longer than ``segment.max_frames`` are
centre-cropped to that length.
"""
from __future__ import annotations
import hashlib
import math
import time
from concurrent.futures import ThreadPoolExecutor
from dataclasses import dataclass, field
from typing import Dict, List, Optional, Sequence, Tuple
import numpy as np
from ..data.annotations import AnnotationError, load_segments
from ..data.availability import Availability
from ..data.io import load_alignment_metadata, load_modality
from ..data.modalities import BENCHMARK_MODALITIES
from ..errors import DataQualityError
from ..paths import DataRoot
from ..taxonomy import Taxonomy
FS = 100 # Hz, every released stream
HEADS = ("verb_fine", "verb_composite", "noun", "hand")
PROTOCOLS = ("sliding", "segment")
# ---------------------------------------------------------------------------
# timeline
# ---------------------------------------------------------------------------
@dataclass(frozen=True)
class Timeline:
rid: str
n_total: int # alignment timeline length (frames, 100 Hz)
offsets: Dict[str, int] # usable benchmark modality -> first timeline frame
lengths: Dict[str, int] # usable benchmark modality -> rows (availability)
start: int # evaluation span start (multiple of ds)
n_frames_ds: int # span length at the decimated rate
ds: int
@property
def end(self) -> int:
return self.start + self.ds * self.n_frames_ds
def _release_window(meta: dict) -> Optional[Tuple[int, int]]:
pads = meta.get("mocap_padding") or {}
pad = pads.get("mocap_sq") or pads.get("mocap") or {}
front, n = pad.get("front_samples"), pad.get("n_frames")
if front is None or n is None:
return None
return int(front), int(n)
def recording_timeline(av: Availability, dr: DataRoot, rid: str, ds: int,
modalities: Sequence[str] = BENCHMARK_MODALITIES) -> Timeline:
rec = av.table["recordings"][rid]
n_total = rec.get("alignment_n_samples") or rec.get("n_samples_100hz")
if not n_total:
raise DataQualityError(f"{rid}: no alignment length in the availability table", "no_length")
n_total = int(n_total)
lengths = {m: int(rec["modalities"][m]["n_frames"]) for m in modalities
if rec["modalities"].get(m, {}).get("status") == "ok"}
if not lengths:
raise DataQualityError(f"{rid}: no usable benchmark modality", "no_modality")
offsets: Dict[str, int] = {}
window = None
for m, L in lengths.items():
if L in (n_total, n_total + 1):
offsets[m] = 0
continue
if window is None:
window = _release_window(load_alignment_metadata(dr, rid)) or (-1, -1)
front, n = window
if L == n and 0 <= front and front + n <= n_total:
offsets[m] = front
else:
raise DataQualityError(
f"{rid}: {m} has {L} rows; the alignment timeline has {n_total} frames and the "
f"release window is {window} -- cannot place the stream", "timeline")
a = max(offsets.values())
b = min(min(offsets[m] + L for m, L in lengths.items()), n_total)
a = int(math.ceil(a / ds) * ds)
n_ds = max(0, (b - a) // ds)
return Timeline(rid, n_total, offsets, lengths, a, n_ds, ds)
# ---------------------------------------------------------------------------
# labels and samples
# ---------------------------------------------------------------------------
@dataclass
class LabelledSpan:
start: int # timeline frame (100 Hz), inclusive
end: int # exclusive
labels: Tuple[int, int, int, int]
def kept_segments(dr: DataRoot, rid: str, tax: Taxonomy, min_sec: float,
counts: Dict[str, int]) -> List[LabelledSpan]:
out: List[LabelledSpan] = []
for s in load_segments(dr, rid):
counts["segments_total"] = counts.get("segments_total", 0) + 1
lab, why = tax.classify(s.verb, s.noun, s.hand)
if lab is None:
key = "dropped_" + why.split(":")[0].replace("dropped_", "")
counts[key] = counts.get(key, 0) + 1
continue
if s.duration_sec < min_sec:
counts["dropped_short"] = counts.get("dropped_short", 0) + 1
continue
out.append(LabelledSpan(int(round(s.start_sec * FS)), int(round(s.end_sec * FS)),
tuple(int(lab[h]) for h in HEADS)))
out.sort(key=lambda x: x.start)
for a, b in zip(out, out[1:]):
if b.start < a.end:
raise AnnotationError(f"{rid}: kept segments overlap ({a.start / FS:.2f}-"
f"{a.end / FS:.2f} s and {b.start / FS:.2f}-{b.end / FS:.2f} s)")
return out
def sliding_samples(tl: Timeline, spans: List[LabelledSpan], window_sec: float,
stride_sec: float, counts: Dict[str, int]):
ds = tl.ds
win = int(round(window_sec * FS / ds))
stride = int(round(stride_sec * FS / ds))
if win < 1 or stride < 1:
raise ValueError("window and stride must be at least one frame at the decimated rate")
starts = np.asarray([s.start for s in spans], dtype=np.int64)
ends = np.asarray([s.end for s in spans], dtype=np.int64)
s0, s1, labs = [], [], []
for w0 in range(0, tl.n_frames_ds - win + 1, stride):
counts["windows_total"] = counts.get("windows_total", 0) + 1
centre = tl.start + ds * (w0 + win // 2)
i = int(np.searchsorted(starts, centre, side="right")) - 1
if i < 0 or centre >= ends[i]:
counts["windows_unlabelled"] = counts.get("windows_unlabelled", 0) + 1
continue
s0.append(w0)
s1.append(w0 + win)
labs.append(spans[i].labels)
return s0, s1, labs
def segment_samples(tl: Timeline, spans: List[LabelledSpan], max_frames: int, min_frames: int,
counts: Dict[str, int]):
ds = tl.ds
s0, s1, labs = [], [], []
for sp in spans:
a = int(math.floor((sp.start - tl.start) / ds + 0.5))
b = int(math.floor((sp.end - tl.start) / ds + 0.5))
if a < 0 or b > tl.n_frames_ds:
counts["dropped_outside_span"] = counts.get("dropped_outside_span", 0) + 1
continue
if b - a < min_frames:
counts["dropped_too_few_frames"] = counts.get("dropped_too_few_frames", 0) + 1
continue
if b - a > max_frames:
c = (a + b) // 2
a = c - max_frames // 2
b = a + max_frames
counts["cropped_long"] = counts.get("cropped_long", 0) + 1
s0.append(a)
s1.append(b)
labs.append(sp.labels)
return s0, s1, labs
# ---------------------------------------------------------------------------
# recordings
# ---------------------------------------------------------------------------
@dataclass
class RecordingSamples:
rid: str
arrays: Dict[str, np.ndarray] # modality -> (n_frames_ds, F) float32
starts: np.ndarray # (N,) decimated frame index into arrays
ends: np.ndarray # (N,) exclusive
labels: np.ndarray # (N, 4) int64 in HEADS order
timeline: Timeline
counts: Dict[str, int] = field(default_factory=dict)
@property
def n(self) -> int:
return int(self.starts.shape[0])
def load_recording_samples(dr: DataRoot, av: Availability, tax: Taxonomy, rid: str,
modalities: Sequence[str], cfg: dict) -> RecordingSamples:
ds = int(cfg["downsample"])
tl = recording_timeline(av, dr, rid, ds)
arrays: Dict[str, np.ndarray] = {}
for m in modalities:
if m not in tl.offsets:
raise DataQualityError(f"{rid}: {m} is not usable according to the availability table "
"but the manifest expects it", "manifest")
ma = load_modality(dr, rid, m)
L = ma.data.shape[0]
if L != tl.lengths[m]:
raise DataQualityError(f"{rid}: {m} has {L} rows, the availability table says "
f"{tl.lengths[m]} (data copy differs from v1.1?)", "rows")
a = tl.start - tl.offsets[m]
arr = ma.data[a:a + ds * tl.n_frames_ds:ds]
if arr.shape[0] != tl.n_frames_ds:
raise DataQualityError(f"{rid}: {m} does not cover the evaluation span", "span")
arrays[m] = np.ascontiguousarray(arr, dtype=np.float32)
counts: Dict[str, int] = {}
spans = kept_segments(dr, rid, tax, float(cfg["segment"]["min_sec"]), counts)
if cfg["protocol"] == "sliding":
sc = cfg["sliding"]
if sc.get("label", "centre") != "centre":
raise ValueError("sliding.label must be 'centre'")
s0, s1, labs = sliding_samples(tl, spans, float(sc["window_sec"]), float(sc["stride_sec"]),
counts)
elif cfg["protocol"] == "segment":
s0, s1, labs = segment_samples(tl, spans, int(cfg["segment"]["max_frames"]),
int(cfg.get("min_frames", 4)), counts)
else:
raise ValueError(f"protocol must be one of {PROTOCOLS}, got {cfg['protocol']!r}")
counts["samples"] = len(s0)
return RecordingSamples(rid, arrays, np.asarray(s0, np.int64), np.asarray(s1, np.int64),
np.asarray(labs, np.int64).reshape(-1, len(HEADS)), tl, counts)
def load_partition(dr: DataRoot, av: Availability, tax: Taxonomy, rids: Sequence[str],
modalities: Sequence[str], cfg: dict, workers: int = 1,
log=print, name: str = "") -> List[RecordingSamples]:
t0 = time.time()
def one(rid):
return load_recording_samples(dr, av, tax, rid, modalities, cfg)
if workers > 1 and len(rids) > 1:
with ThreadPoolExecutor(max_workers=workers) as ex:
recs = list(ex.map(one, rids))
else:
recs = [one(r) for r in rids]
log(f" loaded {name}: {len(recs)} recordings, {sum(r.n for r in recs)} samples "
f"({time.time() - t0:.0f}s)")
return recs
# ---------------------------------------------------------------------------
# normalisation
# ---------------------------------------------------------------------------
def normalisation_stats(recs: Sequence[RecordingSamples], modalities: Sequence[str]
) -> Dict[str, Tuple[np.ndarray, np.ndarray]]:
"""Per-modality mean / SD over the frames of the training samples."""
out = {}
for m in modalities:
s = s2 = None
cnt = 0.0
for r in recs:
if r.n == 0:
continue
T = r.arrays[m].shape[0]
cov = np.zeros(T + 1, dtype=np.float64)
np.add.at(cov, r.starts, 1.0)
np.add.at(cov, r.ends, -1.0)
w = np.cumsum(cov)[:T]
x = r.arrays[m].astype(np.float64)
ws, ws2 = w @ x, w @ (x * x)
s = ws if s is None else s + ws
s2 = ws2 if s2 is None else s2 + ws2
cnt += float(w.sum())
if not cnt:
raise ValueError(f"no training frames for modality {m}")
mu = s / cnt
sd = np.sqrt(np.maximum(s2 / cnt - mu * mu, 0.0))
sd[sd < 1e-8] = 1.0
out[m] = (mu.astype(np.float32), sd.astype(np.float32))
return out
def apply_normalisation(recs: Sequence[RecordingSamples],
stats: Dict[str, Tuple[np.ndarray, np.ndarray]]) -> None:
for r in recs:
for m, (mu, sd) in stats.items():
z = (r.arrays[m] - mu) / sd
r.arrays[m] = np.nan_to_num(z, nan=0.0, posinf=0.0, neginf=0.0).astype(np.float32)
def stats_md5(stats: Dict[str, Tuple[np.ndarray, np.ndarray]]) -> str:
h = hashlib.md5()
for m in sorted(stats):
h.update(m.encode())
h.update(np.ascontiguousarray(stats[m][0]).tobytes())
h.update(np.ascontiguousarray(stats[m][1]).tobytes())
return h.hexdigest()
# ---------------------------------------------------------------------------
# torch dataset
# ---------------------------------------------------------------------------
class SampleSet:
"""Flat view over the samples of several recordings (map-style dataset)."""
def __init__(self, recs: Sequence[RecordingSamples], modalities: Sequence[str],
index: Optional[np.ndarray] = None):
self.recs = list(recs)
self.modalities = list(modalities)
full = np.asarray([(i, j) for i, r in enumerate(self.recs) for j in range(r.n)],
dtype=np.int64).reshape(-1, 2)
self.index = full if index is None else full[np.asarray(index, dtype=np.int64)]
def __len__(self) -> int:
return int(self.index.shape[0])
def __getitem__(self, k: int):
i, j = self.index[k]
r = self.recs[i]
a, b = int(r.starts[j]), int(r.ends[j])
return {m: r.arrays[m][a:b] for m in self.modalities}, r.labels[j], int(i)
@property
def labels(self) -> np.ndarray:
if not len(self):
return np.zeros((0, len(HEADS)), np.int64)
return np.stack([self.recs[i].labels[j] for i, j in self.index])
@property
def rec_index(self) -> np.ndarray:
return self.index[:, 0].copy()
def modality_dims(self) -> Dict[str, int]:
r = self.recs[0]
return {m: int(r.arrays[m].shape[1]) for m in self.modalities}
def collate(batch):
import torch
xs, ys, ri = zip(*batch)
mods = list(xs[0])
lens = np.asarray([x[mods[0]].shape[0] for x in xs], dtype=np.int64)
T = int(lens.max())
out = {}
for m in mods:
F = xs[0][m].shape[1]
buf = np.zeros((len(xs), T, F), dtype=np.float32)
for i, x in enumerate(xs):
buf[i, :x[m].shape[0]] = x[m]
out[m] = torch.from_numpy(buf)
mask = torch.from_numpy(np.arange(T)[None, :] < lens[:, None])
y = torch.from_numpy(np.stack(ys).astype(np.int64))
return out, mask, y, torch.as_tensor(ri, dtype=torch.int64)
def label_counts(labels: np.ndarray, sizes: Dict[str, int]) -> Dict[str, List[int]]:
return {h: np.bincount(labels[:, i], minlength=sizes[h]).astype(int).tolist()
for i, h in enumerate(HEADS)}