Spaces:
Runtime error
Runtime error
File size: 9,132 Bytes
85ba398 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
from typing import Optional
import torch
from fairseq.modules import (
ESPNETMultiHeadedAttention,
LayerNorm,
MultiheadAttention,
RelPositionMultiHeadedAttention,
RotaryPositionMultiHeadedAttention,
)
from fairseq.utils import get_activation_fn
class ConvolutionModule(torch.nn.Module):
"""Convolution block used in the conformer block"""
def __init__(
self,
embed_dim,
channels,
depthwise_kernel_size,
dropout,
activation_fn="swish",
bias=False,
export=False,
):
"""
Args:
embed_dim: Embedding dimension
channels: Number of channels in depthwise conv layers
depthwise_kernel_size: Depthwise conv layer kernel size
dropout: dropout value
activation_fn: Activation function to use after depthwise convolution kernel
bias: If bias should be added to conv layers
export: If layernorm should be exported to jit
"""
super(ConvolutionModule, self).__init__()
assert (
depthwise_kernel_size - 1
) % 2 == 0, "kernel_size should be a odd number for 'SAME' padding"
self.layer_norm = LayerNorm(embed_dim, export=export)
self.pointwise_conv1 = torch.nn.Conv1d(
embed_dim,
2 * channels,
kernel_size=1,
stride=1,
padding=0,
bias=bias,
)
self.glu = torch.nn.GLU(dim=1)
self.depthwise_conv = torch.nn.Conv1d(
channels,
channels,
depthwise_kernel_size,
stride=1,
padding=(depthwise_kernel_size - 1) // 2,
groups=channels,
bias=bias,
)
self.batch_norm = torch.nn.BatchNorm1d(channels)
self.activation = get_activation_fn(activation_fn)(channels)
self.pointwise_conv2 = torch.nn.Conv1d(
channels,
embed_dim,
kernel_size=1,
stride=1,
padding=0,
bias=bias,
)
self.dropout = torch.nn.Dropout(dropout)
def forward(self, x):
"""
Args:
x: Input of shape B X T X C
Returns:
Tensor of shape B X T X C
"""
x = self.layer_norm(x)
# exchange the temporal dimension and the feature dimension
x = x.transpose(1, 2)
# GLU mechanism
x = self.pointwise_conv1(x) # (batch, 2*channel, dim)
x = self.glu(x) # (batch, channel, dim)
# 1D Depthwise Conv
x = self.depthwise_conv(x)
x = self.batch_norm(x)
x = self.activation(x)
x = self.pointwise_conv2(x)
x = self.dropout(x)
return x.transpose(1, 2)
class FeedForwardModule(torch.nn.Module):
"""Positionwise feed forward layer used in conformer"""
def __init__(
self,
input_feat,
hidden_units,
dropout1,
dropout2,
activation_fn="swish",
bias=True,
):
"""
Args:
input_feat: Input feature dimension
hidden_units: Hidden unit dimension
dropout1: dropout value for layer1
dropout2: dropout value for layer2
activation_fn: Name of activation function
bias: If linear layers should have bias
"""
super(FeedForwardModule, self).__init__()
self.layer_norm = LayerNorm(input_feat)
self.w_1 = torch.nn.Linear(input_feat, hidden_units, bias=bias)
self.w_2 = torch.nn.Linear(hidden_units, input_feat, bias=bias)
self.dropout1 = torch.nn.Dropout(dropout1)
self.dropout2 = torch.nn.Dropout(dropout2)
self.activation = get_activation_fn(activation_fn)(hidden_units)
def forward(self, x):
"""
Args:
x: Input Tensor of shape T X B X C
Returns:
Tensor of shape T X B X C
"""
x = self.layer_norm(x)
x = self.w_1(x)
x = self.activation(x)
x = self.dropout1(x)
x = self.w_2(x)
return self.dropout2(x)
class ConformerEncoderLayer(torch.nn.Module):
"""Conformer block based on https://arxiv.org/abs/2005.08100. We currently don't support relative positional encoding in MHA"""
def __init__(
self,
embed_dim,
ffn_embed_dim,
attention_heads,
dropout,
use_fp16,
depthwise_conv_kernel_size=31,
activation_fn="swish",
attn_type=None,
pos_enc_type="abs",
):
"""
Args:
embed_dim: Input embedding dimension
ffn_embed_dim: FFN layer dimension
attention_heads: Number of attention heads in MHA
dropout: dropout value
depthwise_conv_kernel_size: Size of kernel in depthwise conv layer in convolution module
activation_fn: Activation function name to use in convulation block and feed forward block
attn_type: MHA implementation from ESPNET vs fairseq
pos_enc_type: Positional encoding type - abs, rope, rel_pos
"""
self.pos_enc_type = pos_enc_type
super(ConformerEncoderLayer, self).__init__()
self.ffn1 = FeedForwardModule(
embed_dim,
ffn_embed_dim,
dropout,
dropout,
)
self.self_attn_layer_norm = LayerNorm(embed_dim, export=False)
self.self_attn_dropout = torch.nn.Dropout(dropout)
if attn_type == "espnet":
if self.pos_enc_type == "rel_pos":
self.self_attn = RelPositionMultiHeadedAttention(
embed_dim,
attention_heads,
dropout=dropout,
)
elif self.pos_enc_type == "rope":
self.self_attn = RotaryPositionMultiHeadedAttention(
embed_dim, attention_heads, dropout=dropout, precision=use_fp16
)
elif self.pos_enc_type == "abs":
self.self_attn = ESPNETMultiHeadedAttention(
embed_dim,
attention_heads,
dropout=dropout,
)
else:
raise Exception(f"Unsupported attention type {self.pos_enc_type}")
else:
# Default to fairseq MHA
self.self_attn = MultiheadAttention(
embed_dim,
attention_heads,
dropout=dropout,
)
self.conv_module = ConvolutionModule(
embed_dim=embed_dim,
channels=embed_dim,
depthwise_kernel_size=depthwise_conv_kernel_size,
dropout=dropout,
activation_fn=activation_fn,
)
self.ffn2 = FeedForwardModule(
embed_dim,
ffn_embed_dim,
dropout,
dropout,
activation_fn=activation_fn,
)
self.final_layer_norm = LayerNorm(embed_dim, export=False)
def forward(
self,
x,
encoder_padding_mask: Optional[torch.Tensor],
position_emb: Optional[torch.Tensor] = None,
):
"""
Args:
x: Tensor of shape T X B X C
encoder_padding_mask: Optional mask tensor
positions:
Returns:
Tensor of shape T X B X C
"""
residual = x
x = self.ffn1(x)
x = x * 0.5 + residual
residual = x
x = self.self_attn_layer_norm(x)
if self.pos_enc_type == "rel_pos":
x, attn = self.self_attn(
query=x,
key=x,
value=x,
key_padding_mask=encoder_padding_mask,
pos_emb=position_emb,
need_weights=False,
)
else:
x, attn = self.self_attn(
query=x,
key=x,
value=x,
key_padding_mask=encoder_padding_mask,
need_weights=False,
)
x = self.self_attn_dropout(x)
x = x + residual
residual = x
# TBC to BTC
x = x.transpose(0, 1)
x = self.conv_module(x)
# BTC to TBC
x = x.transpose(0, 1)
x = residual + x
residual = x
x = self.ffn2(x)
layer_result = x
x = x * 0.5 + residual
x = self.final_layer_norm(x)
return x, (attn, layer_result)
class ConformerWav2Vec2EncoderLayer(ConformerEncoderLayer):
"""Encoder layer for Wav2vec2 encoder"""
def forward(
self,
x: torch.Tensor,
self_attn_mask: torch.Tensor = None,
self_attn_padding_mask: torch.Tensor = None,
need_weights: bool = False,
att_args=None,
position_emb=None,
):
return super().forward(x, self_attn_padding_mask, position_emb)
|