"""Loading and structural description of pyannote/segmentation-3.0. The checkpoint is a PyTorch Lightning file whose pickle references pyannote classes, so reading it requires pyannote.audio to be importable. Everything downstream of this module works on plain numpy arrays. """ from pathlib import Path import numpy as np REPO = "pyannote/segmentation-3.0" SNAPSHOT = ( Path.home() / ".cache/huggingface/hub/models--pyannote--segmentation-3.0/snapshots" ) SAMPLE_RATE = 16000 DURATION = 10.0 NUM_SAMPLES = 160000 NUM_FRAMES = 589 NUM_CLASSES = 7 HIDDEN = 128 NUM_LAYERS = 4 # SincNet geometry: (kernel_size, stride) for conv, abs, pool, conv, pool, conv, pool. SINC_KERNEL = 251 SINC_STRIDE = 10 SINC_FILTERS = 80 def checkpoint_path() -> Path: """Path of the single snapshot of the model in the local Hugging Face cache.""" snaps = sorted(SNAPSHOT.glob("*/pytorch_model.bin")) if not snaps: raise FileNotFoundError( f"{REPO} not found under {SNAPSHOT}; run \"hf download {REPO}\" first" ) return snaps[-1] def load_state_dict() -> dict: """Returns the checkpoint tensors as numpy arrays keyed by parameter name.""" import torch ckpt = torch.load(checkpoint_path(), map_location="cpu", weights_only=False) return {k: v.numpy().copy() for k, v in ckpt["state_dict"].items()} def sinc_filters(sd: dict) -> np.ndarray: """Materialises the parametric sinc filterbank as a plain [80, 251] matrix. asteroid_filterbanks builds the filters from the learned band edges every forward pass, but the edges are frozen at inference time, so the result is a constant that can be folded into the exported weights. """ low_hz = sd["sincnet.conv1d.0.filterbank.low_hz_"].astype(np.float64) band_hz = sd["sincnet.conv1d.0.filterbank.band_hz_"].astype(np.float64) window = sd["sincnet.conv1d.0.filterbank.window_"].astype(np.float64) n_ = sd["sincnet.conv1d.0.filterbank.n_"].astype(np.float64) min_low_hz, min_band_hz = 50.0, 50.0 low = min_low_hz + np.abs(low_hz) high = np.clip(low + min_band_hz + np.abs(band_hz), min_low_hz, SAMPLE_RATE / 2) band = (high - low)[:, 0] ft_low = low @ n_ ft_high = high @ n_ cos_left = ((np.sin(ft_high) - np.sin(ft_low)) / (n_ / 2)) * window cos = np.concatenate( [cos_left, 2 * band[:, None], np.flip(cos_left, axis=1)], axis=1 ) cos /= 2 * band[:, None] sin_left = ((np.cos(ft_low) - np.cos(ft_high)) / (n_ / 2)) * window sin = np.concatenate( [sin_left, np.zeros_like(band[:, None]), -np.flip(sin_left, axis=1)], axis=1 ) sin /= 2 * band[:, None] return np.concatenate([cos, sin], axis=0).astype(np.float32) def num_frames(num_samples: int) -> int: """Output frame count of the SincNet stack for a given input length.""" geometry = [(SINC_KERNEL, SINC_STRIDE), (3, 3), (5, 1), (3, 3), (5, 1), (3, 3)] n = num_samples for kernel, stride in geometry: n = 1 + (n - kernel) // stride return n def receptive_field() -> tuple[int, int, int]: """(size, step, start) of the output frame grid, in samples. Frame f covers the 991 input samples starting at f * 270. The grid begins exactly at sample 0 because the centre of the first frame sits half a receptive field into the signal. """ geometry = [(SINC_KERNEL, SINC_STRIDE), (3, 3), (5, 1), (3, 3), (5, 1), (3, 3)] def rf_size(frames: int) -> int: size = frames for kernel, stride in reversed(geometry): size = kernel + (size - 1) * stride return size def rf_centre(frame: int) -> int: centre = frame for kernel, stride in reversed(geometry): centre = centre * stride + (kernel - 1) // 2 return centre size = rf_size(1) step = rf_size(2) - size start = rf_centre(0) - (size - 1) // 2 return size, step, start def powerset_mapping(num_classes: int = 3, max_set_size: int = 2) -> np.ndarray: """[num_powerset_classes, num_classes] membership matrix in pyannote order.""" from itertools import combinations rows = [] for size in range(max_set_size + 1): for members in combinations(range(num_classes), size): row = np.zeros(num_classes, dtype=np.float32) row[list(members)] = 1.0 rows.append(row) return np.stack(rows) FRAME_SIZE, FRAME_STEP, FRAME_START = receptive_field() assert num_frames(NUM_SAMPLES) == NUM_FRAMES assert (FRAME_SIZE, FRAME_STEP, FRAME_START) == (991, 270, 0)