Scikit-learn
human-activity-recognition
wearable
wrist
time-series
cpu
scikit-learn
WISP / src /wisp_fast /decoder.py
Zipeng365's picture
wisp-har 0.2.0: accelerated implementation with speed (exact/fast) and HMM sequence (subject/recording) options
ac7f7cb verified
Raw History Blame Contribute Delete
9.92 kB
"""Grouped transition counting and Viterbi decoding on the release class axis.
Two ways to form the sequences that the HMM smoother decodes:
``subject``
All windows of one participant (the ``subject``/``groups`` value) form one
chain, ordered by ``time_index``. This is the original WISP behaviour.
``recording``
Each recording forms its own chain. Within one participant's rows, in the
order they are supplied, a new recording starts whenever ``time_index``
goes backwards. Classifier scores are rescaled by one confidence factor
chosen on validation data (``fit_confidence_scale``), and smoothing is
skipped when most training recordings contain a single activity, because
then there are no within-recording transitions to learn.
"""
from functools import lru_cache
import ctypes
from .native import lookup
import numpy as np
SEQUENCES = ("subject", "recording")
# Smoothing is skipped when more than this fraction of training recordings has one label.
SINGLE_LABEL_RECORDING_LIMIT = 0.5
# Candidate confidence-scaling factors, searched on validation negative log-likelihood.
CONFIDENCE_SCALE_GRID = np.exp(np.linspace(np.log(0.01), np.log(10.0), 121))
def check_sequence(sequence):
if sequence not in SEQUENCES:
raise ValueError(f"sequence must be one of {', '.join(SEQUENCES)}; got {sequence!r}")
return sequence
def recording_ids(groups, time_index):
"""Split each participant into recordings at every backward step of time_index.
Rows keep the order in which they are supplied; the k-th backward step within
a participant starts recording k. Returns string IDs ``"<participant>|<k>"``.
"""
g = np.asarray(groups).astype(str)
if time_index is None:
return g
t = np.asarray(time_index, dtype=float)
if t.shape != g.shape:
raise ValueError("time_index must have one entry per window")
part = np.zeros(len(t), dtype=np.int64)
for subject in np.unique(g):
rows = np.flatnonzero(g == subject)
part[rows] = np.concatenate([[0], np.cumsum(np.diff(t[rows]) < 0)])
return np.char.add(np.char.add(g, "|"), part.astype(str))
def single_label_fraction(y, recordings):
"""Fraction of recordings with at least two windows whose windows all share one label."""
_, inverse, counts = np.unique(recordings, return_inverse=True, return_counts=True)
pairs = np.unique(np.stack([inverse.ravel(), np.asarray(y, dtype=np.int64)]), axis=1)
labels_per_recording = np.bincount(pairs[0], minlength=len(counts))
eligible = counts > 1
if not eligible.any():
return 1.0
return float(1 - np.mean(labels_per_recording[eligible] > 1))
def scale_confidence(proba, scale):
"""Divide log-probabilities by ``scale`` and renormalise (scale 1 leaves them unchanged)."""
P = np.asarray(proba, dtype=np.float64)
if scale == 1.0:
return P
Z = np.log(np.clip(P, 1e-12, 1.0)) / scale
Z -= Z.max(axis=1, keepdims=True)
Q = np.exp(Z)
return Q / Q.sum(axis=1, keepdims=True)
def fit_confidence_scale(proba, y):
"""Choose the factor on CONFIDENCE_SCALE_GRID with the lowest validation NLL."""
L = np.log(np.clip(np.asarray(proba, dtype=np.float64), 1e-12, 1.0))
y = np.asarray(y, dtype=int)
best = (np.inf, 1.0)
for scale in CONFIDENCE_SCALE_GRID:
Z = L / scale
Z -= Z.max(1, keepdims=True)
Q = np.exp(Z)
Q /= Q.sum(1, keepdims=True)
nll = -float(np.mean(np.log(np.clip(Q[np.arange(len(y)), y], 1e-12, 1.0))))
if nll < best[0]:
best = (nll, float(scale))
return best[1]
@lru_cache(maxsize=1)
def _native():
fn = lookup("wisp_viterbi")
if fn is None:
return None
arr = np.ctypeslib.ndpointer
f64 = arr(dtype=np.float64, flags="C_CONTIGUOUS")
i32 = arr(dtype=np.int32, flags="C_CONTIGUOUS")
i64 = arr(dtype=np.int64, flags="C_CONTIGUOUS")
fn.argtypes = [f64, f64, f64, ctypes.c_int64, ctypes.c_int64, i32, f64, f64, i64]
fn.restype = None
return fn
@lru_cache(maxsize=1)
def _native_groups():
fn = lookup("wisp_viterbi_groups")
if fn is not None:
arr = np.ctypeslib.ndpointer
f64 = arr(dtype=np.float64, flags="C_CONTIGUOUS")
i32 = arr(dtype=np.int32, flags="C_CONTIGUOUS")
i64 = arr(dtype=np.int64, flags="C_CONTIGUOUS")
fn.argtypes = [f64, f64, f64, i64, ctypes.c_int64, ctypes.c_int64, i32, f64, f64, i64]
fn.restype = None
return fn
def _order(n, groups, time_index):
"""Preserve original order within a group when timestamps are absent or tied."""
if groups is None:
return np.arange(n), np.array([0, n], dtype=np.intp)
g = np.asarray(groups)
if g.shape != (n,):
raise ValueError("groups must have one entry per window")
if np.any(g != g):
raise ValueError("Missing group ID")
if time_index is None:
order = np.argsort(g, kind="stable")
else:
time = np.asarray(time_index)
if time.shape != (n,):
raise ValueError("time_index must have one entry per window")
order = np.lexsort((time, g))
sorted_groups = g[order]
bounds = np.r_[0, np.flatnonzero(sorted_groups[1:] != sorted_groups[:-1]) + 1, n]
return order, bounds
class FastSmoother:
"""HMM smoother. ``sequence`` selects how windows are grouped into chains (see module docs)."""
def __init__(self, sequence="subject"):
self.sequence = check_sequence(sequence)
@property
def active(self):
"""False when recording-level smoothing was skipped at fit time."""
return getattr(self, "active_", True)
@property
def confidence_scale(self):
return getattr(self, "confidence_scale_", 1.0)
def set_confidence_scale(self, scale):
scale = float(scale)
if not np.isfinite(scale) or scale <= 0:
raise ValueError("confidence scale must be finite and positive")
self.confidence_scale_ = scale
return self
def fit(self, y, *, groups=None, time_index=None, n_classes=None):
y = np.asarray(y, dtype=np.int64)
k = int(n_classes) if n_classes is not None else int(y.max()) + 1
if y.ndim != 1 or y.size == 0 or np.any((y < 0) | (y >= k)):
raise ValueError("Invalid y or n_classes")
if getattr(self, "sequence", "subject") == "recording":
self.confidence_scale_ = 1.0
self.active_ = True
self.single_label_fraction_ = None
if groups is not None:
rec = recording_ids(groups, time_index)
self.single_label_fraction_ = single_label_fraction(y, rec)
self.active_ = self.single_label_fraction_ <= SINGLE_LABEL_RECORDING_LIMIT
groups = rec
A = np.ones((k, k), dtype=np.float64)
if groups is not None:
order, bounds = _order(len(y), groups, time_index)
labels = y[order]
edges = labels[:-1] * k + labels[1:]
valid = np.ones(edges.size, dtype=bool)
valid[bounds[1:-1] - 1] = False
A += np.bincount(edges[valid], minlength=k*k).reshape(k, k)
# Upstream adds persistence twice for ungrouped training and once otherwise.
A.flat[::k + 1] += 4.0 if groups is None else 2.0
p = np.bincount(y, minlength=k).astype(np.float64) + 1.0
self.transition_ = A / A.sum(axis=1, keepdims=True)
self.prior_ = p / p.sum()
self.log_A_ = np.log(np.clip(self.transition_, 1e-12, 1.0))
self.log_pi_ = np.log(np.clip(self.prior_, 1e-12, 1.0))
return self
def predict_from_proba(self, proba, *, groups=None, time_index=None):
P = np.asarray(proba, dtype=np.float64)
if P.ndim != 2 or P.shape[1] != len(self.prior_) or not np.isfinite(P).all():
raise ValueError("Invalid emissions")
n, k = P.shape
if n == 0:
return np.empty(0, dtype=np.int64)
if getattr(self, "sequence", "subject") == "recording":
if not self.active:
return P.argmax(1).astype(np.int64)
P = scale_confidence(P, self.confidence_scale)
if groups is not None:
groups = recording_ids(groups, time_index)
order, bounds = _order(n, groups, time_index)
return self._decode_ordered(P,order,bounds)
def _decode_ordered(self,P,order,bounds):
n,k=P.shape
P = np.clip(P[order], 1e-12, 1.0)
P /= P.sum(axis=1, keepdims=True)
E = np.ascontiguousarray(np.log(np.clip(P, 1e-12, 1.0)))
back = np.empty((np.diff(bounds).max(), k), dtype=np.int32)
prev, nxt = np.empty(k), np.empty(k)
labels = np.empty(n, dtype=np.int64)
grouped = _native_groups() if _native() is not None else None
if grouped is not None:
grouped(E, self.log_A_, self.log_pi_, np.ascontiguousarray(bounds, dtype=np.int64),
len(bounds)-1, k, back, prev, nxt, labels)
out = np.empty_like(labels)
out[order] = labels
return out
fn = _native()
for lo, hi in zip(bounds[:-1], bounds[1:]):
if fn is not None:
fn(E[lo:hi], self.log_A_, self.log_pi_, hi-lo, k, back, prev, nxt, labels[lo:hi])
else:
prev[:] = self.log_pi_ + E[lo]
cols = np.arange(k)
for i in range(lo + 1, hi):
scores = prev[:, None] + self.log_A_
pred = np.argmax(scores, axis=0)
back[i-lo] = pred
prev[:] = scores[pred, cols] + E[i]
labels[hi-1] = np.argmax(prev)
for i in range(hi-2, lo-1, -1):
labels[i] = back[i-lo+1, labels[i+1]]
out = np.empty_like(labels)
out[order] = labels
return out