Zipeng365's picture
Add ICML 2026 TSDecompose benchmark release
17b7ba4 verified
Raw
History Blame Contribute Delete
11.6 kB
# gabor_cluster.py
from __future__ import annotations
import numpy as np
from dataclasses import dataclass
from typing import List, Dict, Optional, Sequence, Tuple
try:
import faiss # faiss-cpu or faiss-gpu
except ImportError as e:
faiss = None
@dataclass
class GaborClusterConfig:
fs: float = 1.0 # sampling freq
win_len: int = 256 # window length L
hop: int = 64 # hop size a
n_fft: Optional[int] = None
window_type: str = "gaussian" # "gaussian" | "hann"
gaussian_sigma: Optional[float] = None
n_clusters: int = 8 # K
max_atoms: int = 200_000 # max TF points to use in training
use_log_amp: bool = True # log(1+|Z|)
random_state: int = 42 # seed for reproducibility
# training iterations for faiss KMeans
n_iter: int = 20
verbose: bool = False
@dataclass
class GaborClusterModel:
"""
Global clustering model learned from many series.
"""
centroids: np.ndarray # (K, d)
mu: np.ndarray # (d,)
sigma: np.ndarray # (d,)
cfg: GaborClusterConfig # Gabor & feature config
def save(self, path: str) -> None:
np.savez_compressed(
path,
centroids=self.centroids.astype(np.float32),
mu=self.mu.astype(np.float32),
sigma=self.sigma.astype(np.float32),
cfg=np.array([self.cfg], dtype=object),
)
@staticmethod
def load(path: str) -> "GaborClusterModel":
data = np.load(path, allow_pickle=True)
centroids = data["centroids"]
mu = data["mu"]
sigma = data["sigma"]
cfg = data["cfg"][0]
return GaborClusterModel(centroids=centroids, mu=mu, sigma=sigma, cfg=cfg)
@dataclass
class DecompResult:
components: Dict[str, np.ndarray]
residual: np.ndarray
meta: Dict
def gabor_components_to_TS(
components: Dict[str, np.ndarray],
model: GaborClusterModel,
trend_freq_thr: float = 0.08,
) -> Dict[str, Optional[np.ndarray]]:
"""
Collapse per-cluster components into trend / seasonal buckets based on centroid frequency.
"""
import re
trend = None
seasonal = None
for key, value in components.items():
match = re.match(r"Cluster_(\d+)$", key)
if not match:
continue
cluster_idx = int(match.group(1))
f_norm = float(model.centroids[cluster_idx, 1])
if f_norm <= trend_freq_thr:
trend = value if trend is None else trend + value
else:
seasonal = value if seasonal is None else seasonal + value
return {"trend": trend, "seasonal": seasonal}
# ---------------- STFT / ISTFT utilities ---------------- #
def _make_window(L: int, wtype: str, sigma: Optional[float]) -> np.ndarray:
if wtype == "gaussian":
if sigma is None:
sigma = L / 6.0
n = np.arange(L) - (L - 1) / 2.0
w = np.exp(-0.5 * (n / sigma) ** 2)
return w / np.sqrt((w ** 2).sum())
elif wtype == "hann":
w = np.hanning(L)
return w / np.sqrt((w ** 2).sum())
else:
raise ValueError(f"Unsupported window_type={wtype}")
def _stft_rfft(x: np.ndarray, L: int, hop: int, n_fft: Optional[int],
window: np.ndarray) -> np.ndarray:
"""
Real-input STFT using rfft. Output shape: (M, K_r), where K_r = n_fft//2 + 1
"""
x = np.asarray(x, dtype=float).ravel()
N = len(x)
if n_fft is None:
n_fft = 1 << int(np.ceil(np.log2(L)))
if N < L:
n_frames = 1
else:
n_frames = 1 + (N - L) // hop
Z = np.empty((n_frames, n_fft // 2 + 1), dtype=np.complex64)
for m in range(n_frames):
start = m * hop
seg = np.zeros(L, dtype=float)
if start + L <= N:
seg[:] = x[start:start + L]
else:
tail = N - start
if tail > 0:
seg[:tail] = x[start:]
segw = seg * window
Z[m, :] = np.fft.rfft(segw, n=n_fft)
return Z
def _istft_rfft(Z: np.ndarray, L: int, hop: int, n_fft: int,
window: np.ndarray, length: int) -> np.ndarray:
"""
Overlap-add ISTFT for rfft coefficients.
Z: (M, K_r), K_r = n_fft//2 + 1
"""
M, K_r = Z.shape
x_rec = np.zeros(length + L, dtype=float)
win_acc = np.zeros(length + L, dtype=float)
for m in range(M):
frame = np.fft.irfft(Z[m, :], n=n_fft).real[:L]
start = m * hop
x_rec[start:start + L] += frame * window
win_acc[start:start + L] += window ** 2
nz = win_acc > 1e-12
x_out = np.zeros_like(x_rec)
x_out[nz] = x_rec[nz] / win_acc[nz]
return x_out[:length]
# ---------------- Feature extraction ---------------- #
def _extract_gabor_features(
x: np.ndarray,
cfg: GaborClusterConfig,
window: np.ndarray
) -> Tuple[np.ndarray, np.ndarray]:
"""
For a single series x, compute STFT and return:
- features: (N_atoms, d) matrix
- Z: complex STFT matrix (M, K_r)
Feature: [t_norm, f_norm, log_amp or amp]
"""
x = np.asarray(x, dtype=float).ravel()
N = len(x)
L = cfg.win_len
hop = cfg.hop
n_fft = cfg.n_fft or (1 << int(np.ceil(np.log2(L))))
Z = _stft_rfft(x, L, hop, n_fft, window) # (M, K_r)
M, K_r = Z.shape
amp = np.abs(Z)
if cfg.use_log_amp:
amp_feat = np.log1p(amp)
else:
amp_feat = amp
# normalized time/freq coordinates
if M > 1:
t_idx = np.linspace(0.0, 1.0, M)
else:
t_idx = np.array([0.0])
if K_r > 1:
f_idx = np.linspace(0.0, 1.0, K_r)
else:
f_idx = np.array([0.0])
T, F = np.meshgrid(t_idx, f_idx, indexing="ij") # (M,K_r)
feats = np.stack(
[
T.ravel().astype(np.float32),
F.ravel().astype(np.float32),
amp_feat.ravel().astype(np.float32),
],
axis=1,
) # (M*K_r, 3)
return feats, Z
# ---------------- FAISS K-means training ---------------- #
def train_gabor_clusters(
series_list: Sequence[np.ndarray],
cfg: GaborClusterConfig,
) -> GaborClusterModel:
"""
Learn global Gabor-atom clusters from a list of 1D series.
Steps:
- For each series: STFT -> [t_norm, f_norm, log_amp] features
- Concatenate across all series
- Subsample up to cfg.max_atoms
- Standardize features
- Run FAISS k-means to get centroids
Returns:
GaborClusterModel
"""
if faiss is None:
raise ImportError(
"faiss is not installed. Please install faiss-cpu or faiss-gpu before "
"using train_gabor_clusters."
)
if len(series_list) == 0:
raise ValueError("series_list is empty.")
L = cfg.win_len
window = _make_window(L, cfg.window_type, cfg.gaussian_sigma)
feat_list = []
for x in series_list:
feats, _ = _extract_gabor_features(x, cfg, window)
feat_list.append(feats)
X = np.concatenate(feat_list, axis=0) # (N_atoms, d)
N_atoms, d = X.shape
if cfg.max_atoms is not None and N_atoms > cfg.max_atoms:
rng = np.random.default_rng(cfg.random_state)
idx = rng.choice(N_atoms, cfg.max_atoms, replace=False)
X = X[idx]
N_atoms = X.shape[0]
# standardize
mu = X.mean(axis=0)
sigma = X.std(axis=0) + 1e-8
X_norm = (X - mu) / sigma
X_norm = X_norm.astype(np.float32)
# FAISS KMeans
k = cfg.n_clusters
if cfg.verbose:
print(f"[GaborCluster] Training FAISS KMeans with K={k}, N={N_atoms}, d={d}")
km = faiss.Kmeans(
d=d,
k=k,
niter=cfg.n_iter,
verbose=cfg.verbose,
seed=cfg.random_state,
)
km.train(X_norm)
centroids = km.centroids # (k, d)
return GaborClusterModel(
centroids=centroids,
mu=mu,
sigma=sigma,
cfg=cfg,
)
# ---------------- Per-series decomposition ---------------- #
def _assign_clusters_faiss(
feats: np.ndarray,
model: GaborClusterModel
) -> np.ndarray:
"""
Assign each feature vector to nearest centroid using FAISS IndexFlatL2.
feats: (N_atoms, d)
Returns: labels (N_atoms,) in [0, K-1]
"""
if faiss is None:
raise ImportError(
"faiss is not installed. Please install faiss-cpu or faiss-gpu before "
"using gabor_cluster_decompose."
)
X = (feats - model.mu) / model.sigma
X = X.astype(np.float32)
d = X.shape[1]
index = faiss.IndexFlatL2(d)
index.add(model.centroids.astype(np.float32))
D, I = index.search(X, 1)
labels = I.ravel()
return labels
def gabor_cluster_decompose(
x: np.ndarray,
model: GaborClusterModel,
max_clusters: Optional[int] = None,
) -> DecompResult:
"""
Decompose a single series x by:
- computing Gabor STFT
- assigning each TF atom to nearest global centroid
- reconstructing each cluster as one component via ISTFT
If max_clusters is not None, only keep the largest-energy clusters and
merge the rest into the residual.
"""
cfg = model.cfg
x = np.asarray(x, dtype=float).ravel()
N = len(x)
L = cfg.win_len
hop = cfg.hop
n_fft = cfg.n_fft or (1 << int(np.ceil(np.log2(L))))
window = _make_window(L, cfg.window_type, cfg.gaussian_sigma)
feats, Z = _extract_gabor_features(x, cfg, window)
labels = _assign_clusters_faiss(feats, model)
M, K_r = Z.shape
K = model.centroids.shape[0]
labels_2d = labels.reshape(M, K_r)
# optional: select top clusters by total energy to keep as components
amp = np.abs(Z)
energy_per_cluster = np.zeros(K, dtype=float)
for j in range(K):
mask = (labels_2d == j)
if np.any(mask):
energy_per_cluster[j] = (amp[mask] ** 2).sum()
if max_clusters is not None and max_clusters < K:
# indices of clusters to keep
keep_idx = np.argsort(energy_per_cluster)[-max_clusters:]
keep_mask = np.zeros(K, dtype=bool)
keep_mask[keep_idx] = True
else:
keep_mask = np.ones(K, dtype=bool)
components: Dict[str, np.ndarray] = {}
used_clusters = []
for j in range(K):
if not keep_mask[j]:
continue
mask = (labels_2d == j).astype(np.float32)
if not np.any(mask):
continue
Zj = Z * mask
xj = _istft_rfft(Zj, L, hop, n_fft, window, N)
components[f"Cluster_{j}"] = xj
used_clusters.append(j)
# residual = x - sum_kept_components
if components:
sum_comp = np.zeros_like(x)
for v in components.values():
sum_comp += v
residual = x - sum_comp
else:
residual = x.copy()
meta = dict(
fs=cfg.fs,
win_len=L,
hop=hop,
n_fft=n_fft,
window_type=cfg.window_type,
gaussian_sigma=cfg.gaussian_sigma,
n_clusters=model.centroids.shape[0],
used_clusters=used_clusters,
max_clusters=max_clusters,
feature_dim=model.centroids.shape[1],
)
return DecompResult(components=components, residual=residual, meta=meta)