feat(model): add support for multi-speaker statistics pooling (#1386)

Co-authored-by: Hervé BREDIN <hbredin@users.noreply.github.com>
This commit is contained in:
Clément PagésandHervé BREDIN authored and GitHub committed 2023-11-14 10:33:56 +01:00
1 parent 2787f123cc
commit 884913e1ac
3 files changed
+205 -28

No files matched your search

+1
View File
@@ -13,6 +13,7 @@
- BREAKING(pipeline): rename `WeSpeakerPretrainedSpeakerEmbedding` to `ONNXWeSpeakerPretrainedSpeakerEmbedding`
- BREAKING(setup): remove `onnxruntime` dependency.
You can still use ONNX `hbredin/wespeaker-voxceleb-resnet34-LM` but you will have to install `onnxruntime` yourself.
- feat(model): add support for multi-speaker statistics pooling
## Version 3.0.1 (2023-09-28)
+73 -28
View File
@@ -1,6 +1,6 @@
# MIT License
#
# Copyright (c) 2020 CNRS
# 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
@@ -26,6 +26,7 @@ from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
class StatsPool(nn.Module):
@@ -40,6 +41,39 @@ class StatsPool(nn.Module):
"""
def _pool(self, sequences: torch.Tensor, weights: torch.Tensor) -> torch.Tensor:
"""Helper function to compute statistics pooling
Assumes that weights are already interpolated to match the number of frames
in sequences and that they encode the activation of only one speaker.
Parameters
----------
sequences : (batch, features, frames) torch.Tensor
Sequences of features.
weights : (batch, frames) torch.Tensor
(Already interpolated) weights.
Returns
-------
output : (batch, 2 * features) torch.Tensor
Concatenation of mean and (unbiased) standard deviation.
"""
weights = weights.unsqueeze(dim=1)
# (batch, 1, frames)
v1 = weights.sum(dim=2) + 1e-8
mean = torch.sum(sequences * weights, dim=2) / v1
dx2 = torch.square(sequences - mean.unsqueeze(2))
v2 = torch.square(weights).sum(dim=2)
var = torch.sum(dx2 * weights, dim=2) / (v1 - v2 / v1)
std = torch.sqrt(var)
return torch.cat([mean, std], dim=1)
def forward(
self, sequences: torch.Tensor, weights: Optional[torch.Tensor] = None
) -> torch.Tensor:
@@ -47,42 +81,53 @@ class StatsPool(nn.Module):
Parameters
----------
sequences : (batch, channel, frames) torch.Tensor
Sequences.
weights : (batch, frames) torch.Tensor, optional
When provided, compute weighted mean and standard deviation.
sequences : (batch, features, frames) torch.Tensor
Sequences of features.
weights : (batch, frames) or (batch, speakers, frames) torch.Tensor, optional
Compute weighted mean and standard deviation, using provided `weights`.
Note
----
`sequences` and `weights` might use a different number of frames, in which case `weights`
are interpolated linearly to reach the number of frames in `sequences`.
Returns
-------
output : (batch, 2 * channel) torch.Tensor
Concatenation of mean and (unbiased) standard deviation.
output : (batch, 2 * features) or (batch, speakers, 2 * features) torch.Tensor
Concatenation of mean and (unbiased) standard deviation. When `weights` are
provided with the `speakers` dimension, `output` is computed for each speaker
separately and returned as (batch, speakers, 2 * channel)-shaped tensor.
"""
if weights is None:
mean = sequences.mean(dim=2)
std = sequences.std(dim=2, unbiased=True)
mean = sequences.mean(dim=-1)
std = sequences.std(dim=-1, correction=1)
return torch.cat([mean, std], dim=-1)
else:
if weights.dim() == 2:
has_speaker_dimension = False
weights = weights.unsqueeze(dim=1)
# (batch, 1, frames)
# (batch, frames) -> (batch, 1, frames)
else:
has_speaker_dimension = True
num_frames = sequences.shape[2]
num_weights = weights.shape[2]
if num_frames != num_weights:
warnings.warn(
f"Mismatch between frames ({num_frames}) and weights ({num_weights}) numbers."
)
weights = F.interpolate(
weights, size=num_frames, mode="linear", align_corners=False
)
# interpolate weights if needed
_, _, num_frames = sequences.shape
_, _, num_weights = weights.shape
if num_frames != num_weights:
warnings.warn(
f"Mismatch between frames ({num_frames}) and weights ({num_weights}) numbers."
)
weights = F.interpolate(
weights, size=num_frames, mode="linear", align_corners=False
)
v1 = weights.sum(dim=2)
mean = torch.sum(sequences * weights, dim=2) / v1
output = rearrange(
torch.vmap(self._pool, in_dims=(None, 1))(sequences, weights),
"speakers batch features -> batch speakers features",
)
dx2 = torch.square(sequences - mean.unsqueeze(2))
v2 = torch.square(weights).sum(dim=2)
if not has_speaker_dimension:
return output.squeeze(dim=1)
var = torch.sum(dx2 * weights, dim=2) / (v1 - v2 / v1)
std = torch.sqrt(var)
return torch.cat([mean, std], dim=1)
return output
+131
View File
@@ -0,0 +1,131 @@
# MIT License
#
# Copyright (c) 2023- 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.
import torch
from pyannote.audio.models.blocks.pooling import StatsPool
def test_stats_pool_weightless():
x = torch.Tensor([[[2.0, 4.0], [2.0, 4.0]], [[1.0, 1.0], [1.0, 1.0]]])
# (batch = 2, features = 2, frames = 2)
stats_pool = StatsPool()
y = stats_pool(x)
# (batch = 2, features = 4)
assert torch.equal(
torch.round(y, decimals=4),
torch.Tensor([[3.0, 3.0, 1.4142, 1.4142], [1.0, 1.0, 0.0, 0.0]]),
)
def test_stats_pool_one_speaker():
x = torch.Tensor([[[2.0, 4.0], [2.0, 4.0]], [[1.0, 1.0], [1.0, 1.0]]])
# (batch = 2, features = 2, frames = 2)
w = torch.Tensor(
[
[0.5, 0.01],
[0.2, 0.1],
]
)
# (batch = 2, frames = 2)
stats_pool = StatsPool()
y = stats_pool(x, weights=w)
# (batch = 2, features = 4)
assert torch.equal(
torch.round(y, decimals=4),
torch.Tensor([[2.0392, 2.0392, 1.4142, 1.4142], [1.0, 1.0, 0.0, 0.0]]),
)
def test_stats_pool_multi_speaker():
x = torch.Tensor([[[2.0, 4.0], [2.0, 4.0]], [[1.0, 1.0], [1.0, 1.0]]])
# (batch = 2, features = 2, frames = 2)
w = torch.Tensor([[[0.1, 0.2], [0.2, 0.3]], [[0.001, 0.001], [0.2, 0.3]]])
# (batch = 2, speakers = 2, frames = 2)
stats_pool = StatsPool()
y = stats_pool(x, weights=w)
# (batch = 2, speakers = 2, features = 4)
assert torch.equal(
torch.round(y, decimals=4),
torch.Tensor(
[
[[3.3333, 3.3333, 1.4142, 1.4142], [3.2, 3.2, 1.4142, 1.4142]],
[[1.0, 1.0, 0.0, 0.0], [1.0, 1.0, 0.0, 0.0]],
]
),
)
def test_stats_pool_frame_mismatch():
x = torch.Tensor([[[2.0, 2.0], [2.0, 2.0]], [[1.0, 1.0], [1.0, 1.0]]])
# (batch = 2, features = 2, frames = 2)
stats_pool = StatsPool()
w = torch.Tensor(
[
[0.5, 0.5, 0.0],
[0.0, 0.5, 0.5],
]
)
# (batch = 2, frames = 3)
y = stats_pool(x, weights=w)
# (batch = 2, features = 4)
assert torch.equal(
torch.round(y, decimals=4),
torch.Tensor([[2.0, 2.0, 0.0, 0.0], [1.0, 1.0, 0.0, 0.0]]),
)
def test_stats_pool_all_zero_weights():
x = torch.Tensor([[[2.0, 4.0], [2.0, 4.0]], [[1.0, 1.0], [1.0, 1.0]]])
# (batch = 2, features = 2, frames = 2)
w = torch.Tensor(
[
[0.5, 0.01],
[0.0, 0.0], # all zero weights
]
)
# (batch = 2, frames = 2)
stats_pool = StatsPool()
y = stats_pool(x, weights=w)
# (batch = 2, features = 4)
assert torch.equal(
torch.round(y, decimals=4),
torch.Tensor([[2.0392, 2.0392, 1.4142, 1.4142], [0.0, 0.0, 0.0, 0.0]]),
)