mirror of
https://github.com/pyannote/pyannote-audio.git
synced 2026-11-20 13:07:51 +09:00
Merge branch 'develop' of github.com:pyannote/pyannote-audio into develop
This commit is contained in:
commit
c0125c7c06
2 files changed
+78
-34
No files matched your search
@@ -25,8 +25,10 @@
|
||||
|
||||
# AUTHORS
|
||||
# Hervé BREDIN - http://herve.niderb.fr
|
||||
# Juan Manuel CORIA
|
||||
# Juan Manuel CORIA - https://juanmc2005.github.io
|
||||
|
||||
import math
|
||||
import warnings
|
||||
import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
@@ -34,7 +36,7 @@ from .classification import Classification
|
||||
|
||||
|
||||
class ArcLinear(nn.Module):
|
||||
"""Additive Angular Margin linear module (ArcFace)
|
||||
"""Additive Angular Margin classification module
|
||||
|
||||
Parameters
|
||||
----------
|
||||
@@ -44,18 +46,18 @@ class ArcLinear(nn.Module):
|
||||
Number of classes
|
||||
margin : float
|
||||
Angular margin to penalize distances between embeddings and centers
|
||||
s : float
|
||||
scale : float
|
||||
Scaling factor for the logits
|
||||
"""
|
||||
|
||||
def __init__(self, nfeat, nclass, margin, s):
|
||||
def __init__(self, nfeat, nclass, margin, scale):
|
||||
super(ArcLinear, self).__init__()
|
||||
eps = 1e-4
|
||||
self.min_cos = eps - 1
|
||||
self.max_cos = 1 - eps
|
||||
self.nclass = nclass
|
||||
self.margin = margin
|
||||
self.s = s
|
||||
self.scale = scale
|
||||
self.W = nn.Parameter(torch.Tensor(nclass, nfeat))
|
||||
nn.init.xavier_uniform_(self.W)
|
||||
|
||||
@@ -92,13 +94,16 @@ class ArcLinear(nn.Module):
|
||||
one_hot = torch.zeros_like(cos_theta_j)
|
||||
one_hot.scatter_(1, target, 1.0)
|
||||
# project margin differences into cosθj
|
||||
return self.s * (cos_theta_j + one_hot * (cos_theta_yi_margin - cos_theta_yi))
|
||||
return self.scale * (cos_theta_j + one_hot * (cos_theta_yi_margin - cos_theta_yi))
|
||||
|
||||
|
||||
class AdditiveAngularMarginLoss(Classification):
|
||||
"""Additive angular margin loss
|
||||
|
||||
TODO explain
|
||||
Penalize the distance between an example and its class center
|
||||
with an angular margin, applying a scaling parameter to all logits.
|
||||
The intuition is to bring examples from a single class closer together
|
||||
in the embedding space by means of the angular margin.
|
||||
|
||||
Parameters
|
||||
----------
|
||||
@@ -124,13 +129,13 @@ class AdditiveAngularMarginLoss(Classification):
|
||||
Defaults to 0 (i.e. keep them all).
|
||||
margin : float, optional
|
||||
Angular margin value. Defaults to 0.1.
|
||||
s : float, optional
|
||||
Scaling parameter value for the logits. Defaults to 7.
|
||||
scale : float, optional
|
||||
Scaling parameter value for the logits. Defaults to sqrt(2) * log(n_classes - 1).
|
||||
|
||||
Reference
|
||||
---------
|
||||
TODO
|
||||
|
||||
ArcFace: Additive Angular Margin Loss for Deep Face Recognition
|
||||
http://openaccess.thecvf.com/content_CVPR_2019/html/Deng_ArcFace_Additive_Angular_Margin_Loss_for_Deep_Face_Recognition_CVPR_2019_paper.html
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
@@ -143,7 +148,9 @@ class AdditiveAngularMarginLoss(Classification):
|
||||
per_epoch: float = None,
|
||||
label_min_duration: float = 0.0,
|
||||
margin: float = 0.1,
|
||||
s: float = 7.0,
|
||||
# `s` is deprecated in favor of `scale`
|
||||
s: float = None,
|
||||
scale: float = None
|
||||
):
|
||||
|
||||
super().__init__(
|
||||
@@ -155,8 +162,15 @@ class AdditiveAngularMarginLoss(Classification):
|
||||
per_epoch=per_epoch,
|
||||
label_min_duration=label_min_duration,
|
||||
)
|
||||
|
||||
self.margin = margin
|
||||
self.s = s
|
||||
if s is not None:
|
||||
msg = "The 's' parameter is deprecated in favor of 'scale', " \
|
||||
"and will be removed in a future release"
|
||||
warnings.warn(msg, FutureWarning)
|
||||
self.scale = s
|
||||
else:
|
||||
self.scale = scale
|
||||
|
||||
def more_parameters(self):
|
||||
"""Initialize classifier layer
|
||||
@@ -167,11 +181,16 @@ class AdditiveAngularMarginLoss(Classification):
|
||||
Parameters
|
||||
"""
|
||||
|
||||
nclass = len(self.specifications["y"]["classes"])
|
||||
# Use scaling initialization trick from AdaCos
|
||||
# Reference: https://arxiv.org/abs/1905.00292
|
||||
scale = math.sqrt(2) * math.log(nclass - 1) if self.scale is None else self.scale
|
||||
|
||||
self.classifier_ = ArcLinear(
|
||||
self.model.dimension,
|
||||
len(self.specifications["y"]["classes"]),
|
||||
nclass,
|
||||
self.margin,
|
||||
self.s,
|
||||
scale,
|
||||
).to(self.device)
|
||||
|
||||
return self.classifier_.parameters()
|
||||
@@ -25,8 +25,10 @@
|
||||
|
||||
# AUTHORS
|
||||
# Hervé BREDIN - http://herve.niderb.fr
|
||||
# Juan Manuel CORIA
|
||||
# Juan Manuel CORIA - https://juanmc2005.github.io
|
||||
|
||||
import math
|
||||
import warnings
|
||||
import torch
|
||||
from torch import nn
|
||||
import torch.nn.functional as F
|
||||
@@ -36,19 +38,19 @@ from .classification import Classification
|
||||
class CocoLinear(nn.Module):
|
||||
"""Congenerous Cosine linear module (for CoCo loss)
|
||||
|
||||
Parameters
|
||||
----------
|
||||
nfeat : int
|
||||
Embedding dimension
|
||||
nclass : int
|
||||
Number of classes
|
||||
alpha : float
|
||||
Scaling factor used in embedding L2-normalization
|
||||
"""
|
||||
Parameters
|
||||
----------
|
||||
nfeat : int
|
||||
Embedding dimension
|
||||
nclass : int
|
||||
Number of classes
|
||||
scale : float
|
||||
Scaling factor used in embedding L2-normalization
|
||||
"""
|
||||
|
||||
def __init__(self, nfeat, nclass, alpha):
|
||||
def __init__(self, nfeat, nclass, scale):
|
||||
super(CocoLinear, self).__init__()
|
||||
self.alpha = alpha
|
||||
self.scale = scale
|
||||
self.centers = nn.Parameter(torch.randn(nclass, nfeat))
|
||||
|
||||
def forward(self, x, target=None):
|
||||
@@ -67,14 +69,19 @@ class CocoLinear(nn.Module):
|
||||
# normalize centers
|
||||
cnorm = F.normalize(self.centers)
|
||||
# normalize scaled embeddings
|
||||
xnorm = self.alpha * F.normalize(x)
|
||||
xnorm = self.scale * F.normalize(x)
|
||||
# calculate logits like in `nn.Linear`
|
||||
logits = torch.matmul(xnorm, torch.transpose(cnorm, 0, 1))
|
||||
return logits
|
||||
|
||||
|
||||
class CongenerousCosineLoss(Classification):
|
||||
"""Train embeddings as last hidden layer of an additive angular margin loss classifier
|
||||
"""Congenerous cosine loss
|
||||
|
||||
Train embeddings by maximizing the cosine similarity between
|
||||
an embedding and its class center.
|
||||
A hyper-parameter `alpha` is used to scale logits before
|
||||
applying softmax.
|
||||
|
||||
Parameters
|
||||
----------
|
||||
@@ -98,8 +105,13 @@ class CongenerousCosineLoss(Classification):
|
||||
label_min_duration : `float`, optional
|
||||
Remove speakers with less than that many seconds of speech.
|
||||
Defaults to 0 (i.e. keep them all).
|
||||
alpha : float
|
||||
Scaling factor used in embedding L2-normalization. Defaults to 6.25.
|
||||
scale : float
|
||||
Scaling factor used in embedding L2-normalization. Defaults to sqrt(2) * log(n_classes - 1).
|
||||
|
||||
Reference
|
||||
---------
|
||||
Rethinking Feature Discrimination and Polymerization for Large-scale Recognition
|
||||
https://arxiv.org/abs/1710.00870
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
@@ -111,7 +123,9 @@ class CongenerousCosineLoss(Classification):
|
||||
per_fold: int = 32,
|
||||
per_epoch: float = None,
|
||||
label_min_duration: float = 0.0,
|
||||
alpha: float = 6.25,
|
||||
# `alpha` is deprecated in favor of `scale`
|
||||
alpha: float = None,
|
||||
scale: float = None
|
||||
):
|
||||
|
||||
super().__init__(
|
||||
@@ -124,7 +138,13 @@ class CongenerousCosineLoss(Classification):
|
||||
label_min_duration=label_min_duration,
|
||||
)
|
||||
|
||||
self.alpha = alpha
|
||||
if alpha is not None:
|
||||
msg = "The 'alpha' parameter is deprecated in favor of 'scale', " \
|
||||
"and will be removed in a future release"
|
||||
warnings.warn(msg, FutureWarning)
|
||||
self.scale = alpha
|
||||
else:
|
||||
self.scale = scale
|
||||
|
||||
def more_parameters(self):
|
||||
"""Initialize trainable trainer parameters
|
||||
@@ -135,8 +155,13 @@ class CongenerousCosineLoss(Classification):
|
||||
Trainable trainer parameters
|
||||
"""
|
||||
|
||||
nclass = len(self.specifications["y"]["classes"])
|
||||
# Use scaling initialization trick from AdaCos
|
||||
# Reference: https://arxiv.org/abs/1905.00292
|
||||
scale = math.sqrt(2) * math.log(nclass - 1) if self.scale is None else self.scale
|
||||
|
||||
self.classifier_ = CocoLinear(
|
||||
self.model.dimension, len(self.specifications["y"]["classes"]), self.alpha
|
||||
self.model.dimension, nclass, scale
|
||||
).to(self.device)
|
||||
|
||||
return self.classifier_.parameters()
|
||||
Reference in new issue
Block a user