File size: 10,516 Bytes
7ef8c9b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 | from dataclasses import dataclass
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import Wav2Vec2Model, PreTrainedModel
from transformers.modeling_outputs import ModelOutput
from .ge2e_loss import GE2ELoss
from .speaker_embedding_config import SpeakerEmbeddingConfig
@dataclass
class EmbeddingOutput(ModelOutput):
loss: Optional[torch.FloatTensor] = None
embeddings: torch.Tensor = None # shape (batch_size, embedding_size)
dvecs: Optional[torch.Tensor] = None # shape (N, M, embedding_size) for analysis
# https://arxiv.org/pdf/1803.10963
class AttentiveStatisticsPooling(nn.Module):
def __init__(self, input_dim, hidden_dim=128):
super().__init__()
self.attention = nn.Sequential(
nn.Linear(input_dim, hidden_dim),
nn.Tanh(),
nn.Linear(hidden_dim, 1)
)
def forward(self, x, mask=None):
# x: (batch, time, features)
# Compute attention weights
attn_weights = self.attention(x) # (batch, time, 1)
if mask is not None:
# mask: (batch, time), True = valid, False = padding
attn_weights = attn_weights.masked_fill(~mask.unsqueeze(-1), float('-inf'))
attn_weights = F.softmax(attn_weights, dim=1) # (batch, time, 1)
# Weighted mean
mean = torch.sum(x * attn_weights, dim=1) # (batch, features)
# Weighted std
variance = torch.sum(attn_weights * (x - mean.unsqueeze(1)) ** 2, dim=1)
std = torch.sqrt(variance.clamp(min=1e-5))
# Concatenate mean and std
return torch.cat([mean, std], dim=-1) # (batch, features * 2)
class ZeroModule(nn.Module):
def forward(self, hidden_states):
return torch.zeros_like(hidden_states)
class SpeakerEmbeddingArchitecture(PreTrainedModel):
config_class = SpeakerEmbeddingConfig
def __init__(self, config):
super().__init__(config)
self.config = config
self.encoder = Wav2Vec2Model(config)
self.encoder.init_weights()
if getattr(config, 'disable_positional_embeddings', False):
self.encoder.encoder.pos_conv_embed = ZeroModule()
# Projection layer to get desired embedding dimension
self.pooling = AttentiveStatisticsPooling(config.hidden_size, hidden_dim=config.hidden_size)
self.embedding_size = getattr(config, 'embedding_size', 256)
n_projection_layers = getattr(config, 'n_projection_layers', 1)
self.projection_layers = nn.ModuleList([])
for l in range(n_projection_layers - 1):
self.projection_layers.append(
nn.Sequential(
nn.Linear(
config.hidden_size * 2 if l == 0 else self.embedding_size,
self.embedding_size
),
#nn.BatchNorm1d(self.embedding_size, momentum=0.01, eps=1e-5),
nn.LeakyReLU(0.01)
)
)
self.final_projection = nn.Sequential(
nn.Linear(
config.hidden_size * 2 if n_projection_layers == 1 else self.embedding_size,
self.embedding_size
),
#nn.BatchNorm1d(self.embedding_size, momentum=0.01, eps=1e-5),
nn.LeakyReLU(0.01)
)
self.layer_weights = None
if getattr(config, 'use_layer_weights', False):
num_layers = config.num_hidden_layers + 1
self.layer_weights = nn.Parameter(torch.ones(num_layers), requires_grad=True)
# Optional: Layer normalization before pooling
self.layer_norm = nn.LayerNorm(config.hidden_size)
self.margin = config.loss_margin
self.scale = config.loss_scale
for m in self.pooling.attention.modules():
if isinstance(m, nn.Linear):
nn.init.xavier_uniform_(m.weight, gain=0.2)
if m.bias is not None:
nn.init.zeros_(m.bias)
# Normal init for projection layers
self.projection_layers.apply(self._init_weights)
self.final_projection.apply(self._init_weights)
# GE2E Loss
# self.ge2e_loss = GE2ELoss(loss_method='softmax')
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, nn.Linear):
nn.init.xavier_uniform_(module.weight, gain=0.5)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.BatchNorm1d):
nn.init.ones_(module.weight)
nn.init.zeros_(module.bias)
def temporal_pooling(self, hidden_states, attention_mask=None):
"""
Pool over time dimension with optional attention mask.
Args:
hidden_states: (batch_size, seq_len, hidden_size)
attention_mask: (batch_size, seq_len) - optional
Returns:
pooled: (batch_size, hidden_size)
"""
if attention_mask is not None:
# Mask padding tokens before pooling
# attention_mask shape: (batch_size, seq_len)
mask_expanded = attention_mask.unsqueeze(-1).expand(hidden_states.size())
sum_hidden = torch.sum(hidden_states * mask_expanded, dim=1)
sum_mask = torch.clamp(mask_expanded.sum(dim=1), min=1e-9)
pooled = sum_hidden / sum_mask
else:
# Simple mean pooling
pooled = torch.mean(hidden_states, dim=1)
return pooled
def forward(
self,
input_values: Optional[torch.Tensor],
attention_mask: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
mask_time_indices: Optional[torch.FloatTensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
cached_centroids: Optional[torch.Tensor] = None,
cached_labels: Optional[torch.Tensor] = None,
):
"""
Args:
input_values: (N*M, seq_len) audio input
attention_mask: (N*M, seq_len) optional
compute_loss: whether to compute GE2E loss (needs N and M)
speakers_per_batch: N - number of speakers in batch
utterances_per_speaker: M - utterances per speaker
Returns:
EmbeddingOutput with loss and embeddings
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# 1. Encode audio
encoder_outputs = self.encoder(
input_values=input_values,
attention_mask=attention_mask,
mask_time_indices=mask_time_indices,
output_attentions=output_attentions,
output_hidden_states=True,
return_dict=return_dict,
)
if self.layer_weights is None:
hidden_states = encoder_outputs.last_hidden_state
else:
# Weighted sum of hidden states from all layers
hidden_states = torch.stack(encoder_outputs.hidden_states, dim=0)
layer_weights = F.softmax(self.layer_weights, dim=0)
hidden_states = (hidden_states * layer_weights.view(-1, 1, 1, 1)).sum(dim=0)
# hidden_states = encoder_outputs.last_hidden_state # (N*M, seq_len, hidden_size)
sub_attention_mask = None
if attention_mask is not None:
# Use the built-in Wav2Vec2 function!
sub_attention_mask = self.encoder._get_feature_vector_attention_mask(
feature_vector_length=hidden_states.shape[1],
attention_mask=attention_mask,
add_adapter=False
)
# 2. Optional normalization
hidden_states = self.layer_norm(hidden_states)
# 3. Temporal pooling
pooled = self.pooling(hidden_states, mask=sub_attention_mask)
# pooled = self.temporal_pooling(hidden_states, sub_attention_mask) # (N*M, hidden_size)
# 4. Project to embedding space
# embeddings = self.projection(pooled) # (N*M, embedding_size)
embeddings = pooled
for projection_layer in self.projection_layers:
embeddings = projection_layer(embeddings)
embeddings = self.final_projection(embeddings)
# 5. L2 normalize embeddings (important for GE2E and cosine similarity!)
embeddings = nn.functional.normalize(embeddings, p=2, dim=1, eps=1e-8)
if cached_centroids is not None:
# We are inside the DDP Forward pass now!
# DDP will see 'w' and 'b' being used here.
# w = self.ge2e_loss.w
# b = self.ge2e_loss.b
# Calculate Similarity against CACHED centroids
# embeddings: (Batch, D), cached_centroids: (N, D)
# sim_matrix = torch.mm(embeddings, cached_centroids.transpose(0, 1))
sim_matrix = torch.einsum('bd,bnd->bn', embeddings, cached_centroids)
# sim_matrix = torch.clamp(sim_matrix, min=1e-6)
# sim_matrix = w * sim_matrix + b
# https://arxiv.org/pdf/1801.05599 this loss here
one_hot = torch.zeros_like(sim_matrix)
one_hot.scatter_(1, cached_labels.unsqueeze(1), 1.0)
sim_matrix = sim_matrix - one_hot * self.margin
sim_matrix = sim_matrix * self.scale
if cached_labels is not None:
# loss_fct = nn.CrossEntropyLoss()
loss = F.cross_entropy(sim_matrix, cached_labels)
# Return the Loss directly
return EmbeddingOutput(loss=loss, embeddings=embeddings)
else:
return EmbeddingOutput(loss=None, embeddings=embeddings)
return EmbeddingOutput(loss=None, embeddings=embeddings)
def get_speaker_embedding(self, audio, attention_mask=None):
"""
Convenience method to get embedding for a single audio sample.
Args:
audio: (1, seq_len) or (seq_len,)
attention_mask: optional
Returns:
embedding: (embedding_size,)
"""
if audio.dim() == 1:
audio = audio.unsqueeze(0)
output = self.forward(
input_values=audio,
attention_mask=attention_mask,
compute_loss=False,
return_dict=True
)
return output.embeddings.squeeze(0)
|