Files
pyannote-audio/src/pyannote/audio/models/segmentation/PyanNet.py
T
ea18bc70d8 Align audio CI and releases; require Python 3.11 (#2081)
* feat(cli): add --subfolder option to benchmark command (#2063)
* fix(utils): fix `SpeakerDiarizationMixin.optimal_mapping` when `reference` is a mapping
* fix(task): fix `balance` option crashing training with `KeyError: 'metadata'` (#2067)
* fix(io): fix `Audio.crop` raising when an in-memory waveform is cropped up to its very end (#2068)
* fix(model): make Model.default_activation return identity for unsupported tasks
* fix: compute sample duration from preloaded waveform
* fix(pipeline): omit unset runtime options from model references (#2080)

Co-authored-by: abeeranajam31 <128603901+abeeranajam31@users.noreply.github.com>
Co-authored-by: antoinelaurent <antoinelaurent@me.com>
Co-authored-by: Kayvan Zahiri <123409429+Kayvan-Zahiri@users.noreply.github.com>
Co-authored-by: Mohammad Hijjawi <51020382+MohammadHijjawi97@users.noreply.github.com>
2026-10-22 15:07:01 +02:00

241 lines
7.8 KiB
Python

# MIT License
#
# Copyright (c) 2020 CNRS
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
from functools import lru_cache
from typing import ClassVar
import torch
import torch.nn.functional as F
from einops import rearrange
from pyannote.core.utils.generators import pairwise
from torch import nn
from pyannote.audio.core.model import Model
from pyannote.audio.core.task import Task
from pyannote.audio.models.blocks.sincnet import SincNet
from pyannote.audio.utils.params import merge_dict
class PyanNet(Model):
"""PyanNet segmentation model
SincNet > LSTM > Feed forward > Classifier
Parameters
----------
sample_rate : int, optional
Audio sample rate. Defaults to 16kHz (16000).
num_channels : int, optional
Number of channels. Defaults to mono (1).
sincnet : dict, optional
Keyword arguments passed to the SincNet block.
Defaults to {"stride": 1}.
lstm : dict, optional
Keyword arguments passed to the LSTM layer.
Defaults to {"hidden_size": 128, "num_layers": 2, "bidirectional": True},
i.e. two bidirectional layers with 128 units each.
Set "monolithic" to False to split monolithic multi-layer LSTM into multiple mono-layer LSTMs.
This may prove useful for probing LSTM internals.
linear : dict, optional
Keyword arguments used to initialize linear layers
Defaults to {"hidden_size": 128, "num_layers": 2},
i.e. two linear layers with 128 units each.
"""
SINCNET_DEFAULTS: ClassVar[dict] = {"stride": 10}
LSTM_DEFAULTS: ClassVar[dict] = {
"hidden_size": 128,
"num_layers": 2,
"bidirectional": True,
"monolithic": True,
"dropout": 0.0,
}
LINEAR_DEFAULTS: ClassVar[dict] = {"hidden_size": 128, "num_layers": 2}
def __init__(
self,
sincnet: dict | None = None,
lstm: dict | None = None,
linear: dict | None = None,
sample_rate: int = 16000,
num_channels: int = 1,
task: Task | None = None,
):
super().__init__(sample_rate=sample_rate, num_channels=num_channels, task=task)
sincnet = merge_dict(self.SINCNET_DEFAULTS, sincnet)
sincnet["sample_rate"] = sample_rate
lstm = merge_dict(self.LSTM_DEFAULTS, lstm)
lstm["batch_first"] = True
linear = merge_dict(self.LINEAR_DEFAULTS, linear)
self.save_hyperparameters("sincnet", "lstm", "linear")
self.sincnet = SincNet(**self.hparams.sincnet)
monolithic = lstm["monolithic"]
if monolithic:
multi_layer_lstm = dict(lstm)
del multi_layer_lstm["monolithic"]
self.lstm = nn.LSTM(60, **multi_layer_lstm)
else:
num_layers = lstm["num_layers"]
if num_layers > 1:
self.dropout = nn.Dropout(p=lstm["dropout"])
one_layer_lstm = dict(lstm)
one_layer_lstm["num_layers"] = 1
one_layer_lstm["dropout"] = 0.0
del one_layer_lstm["monolithic"]
self.lstm = nn.ModuleList(
[
nn.LSTM(
60
if i == 0
else lstm["hidden_size"] * (2 if lstm["bidirectional"] else 1),
**one_layer_lstm,
)
for i in range(num_layers)
]
)
if linear["num_layers"] < 1:
return
lstm_out_features: int = self.hparams.lstm["hidden_size"] * (
2 if self.hparams.lstm["bidirectional"] else 1
)
self.linear = nn.ModuleList(
[
nn.Linear(in_features, out_features)
for in_features, out_features in pairwise(
[
lstm_out_features,
]
+ [self.hparams.linear["hidden_size"]]
* self.hparams.linear["num_layers"]
)
]
)
@property
def dimension(self) -> int:
"""Dimension of output"""
if isinstance(self.specifications, tuple):
raise ValueError("PyanNet does not support multi-tasking.")
if self.specifications.powerset:
return self.specifications.num_powerset_classes
else:
return len(self.specifications.classes)
def build(self):
if self.hparams.linear["num_layers"] > 0:
in_features = self.hparams.linear["hidden_size"]
else:
in_features = self.hparams.lstm["hidden_size"] * (
2 if self.hparams.lstm["bidirectional"] else 1
)
self.classifier = nn.Linear(in_features, self.dimension)
self.activation = self.default_activation()
@lru_cache
def num_frames(self, num_samples: int) -> int:
"""Compute number of output frames for a given number of input samples
Parameters
----------
num_samples : int
Number of input samples
Returns
-------
num_frames : int
Number of output frames
"""
return self.sincnet.num_frames(num_samples)
def receptive_field_size(self, num_frames: int = 1) -> int:
"""Compute size of receptive field
Parameters
----------
num_frames : int, optional
Number of frames in the output signal
Returns
-------
receptive_field_size : int
Receptive field size.
"""
return self.sincnet.receptive_field_size(num_frames=num_frames)
def receptive_field_center(self, frame: int = 0) -> int:
"""Compute center of receptive field
Parameters
----------
frame : int, optional
Frame index
Returns
-------
receptive_field_center : int
Index of receptive field center.
"""
return self.sincnet.receptive_field_center(frame=frame)
def forward(self, waveforms: torch.Tensor) -> torch.Tensor:
"""Pass forward
Parameters
----------
waveforms : (batch, channel, sample)
Returns
-------
scores : (batch, frame, classes)
"""
outputs = self.sincnet(waveforms)
if self.hparams.lstm["monolithic"]:
outputs, _ = self.lstm(
rearrange(outputs, "batch feature frame -> batch frame feature")
)
else:
outputs = rearrange(outputs, "batch feature frame -> batch frame feature")
for i, lstm in enumerate(self.lstm):
outputs, _ = lstm(outputs)
if i + 1 < self.hparams.lstm["num_layers"]:
outputs = self.dropout(outputs)
if self.hparams.linear["num_layers"] > 0:
for linear in self.linear:
outputs = F.leaky_relu(linear(outputs))
return self.activation(self.classifier(outputs))