File size: 10,468 Bytes
167f5d0 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 | from typing import Tuple, Dict
import copy
import os
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from funasr.register import tables
class LinearTransform(nn.Module):
def __init__(self, input_dim, output_dim):
super(LinearTransform, self).__init__()
self.input_dim = input_dim
self.output_dim = output_dim
self.linear = nn.Linear(input_dim, output_dim, bias=False)
def forward(self, input):
output = self.linear(input)
return output
class AffineTransform(nn.Module):
def __init__(self, input_dim, output_dim):
super(AffineTransform, self).__init__()
self.input_dim = input_dim
self.output_dim = output_dim
self.linear = nn.Linear(input_dim, output_dim)
def forward(self, input):
output = self.linear(input)
return output
class RectifiedLinear(nn.Module):
def __init__(self, input_dim, output_dim):
super(RectifiedLinear, self).__init__()
self.dim = input_dim
self.relu = nn.ReLU()
self.dropout = nn.Dropout(0.1)
def forward(self, input):
out = self.relu(input)
return out
class FSMNBlock(nn.Module):
def __init__(
self,
input_dim: int,
output_dim: int,
lorder=None,
rorder=None,
lstride=1,
rstride=1,
):
super(FSMNBlock, self).__init__()
self.dim = input_dim
if lorder is None:
return
self.lorder = lorder
self.rorder = rorder
self.lstride = lstride
self.rstride = rstride
self.conv_left = nn.Conv2d(
self.dim, self.dim, [lorder, 1], dilation=[lstride, 1], groups=self.dim, bias=False
)
if self.rorder > 0:
self.conv_right = nn.Conv2d(
self.dim, self.dim, [rorder, 1], dilation=[rstride, 1], groups=self.dim, bias=False
)
else:
self.conv_right = None
def forward(self, input: torch.Tensor, cache: torch.Tensor = None):
x = torch.unsqueeze(input, 1)
x_per = x.permute(0, 3, 2, 1) # B D T C
if cache is not None:
cache = cache.to(x_per.device)
y_left = torch.cat((cache, x_per), dim=2)
cache = y_left[:, :, -(self.lorder - 1) * self.lstride :, :]
else:
y_left = F.pad(x_per, [0, 0, (self.lorder - 1) * self.lstride, 0])
y_left = self.conv_left(y_left)
out = x_per + y_left
if self.conv_right is not None:
# maybe need to check
y_right = F.pad(x_per, [0, 0, 0, self.rorder * self.rstride])
y_right = y_right[:, :, self.rstride :, :]
y_right = self.conv_right(y_right)
out += y_right
out_per = out.permute(0, 3, 2, 1)
output = out_per.squeeze(1)
return output, cache
class BasicBlock(nn.Module):
def __init__(
self,
linear_dim: int,
proj_dim: int,
lorder: int,
rorder: int,
lstride: int,
rstride: int,
stack_layer: int,
):
super(BasicBlock, self).__init__()
self.lorder = lorder
self.rorder = rorder
self.lstride = lstride
self.rstride = rstride
self.stack_layer = stack_layer
self.linear = LinearTransform(linear_dim, proj_dim)
self.fsmn_block = FSMNBlock(proj_dim, proj_dim, lorder, rorder, lstride, rstride)
self.affine = AffineTransform(proj_dim, linear_dim)
self.relu = RectifiedLinear(linear_dim, linear_dim)
def forward(self, input: torch.Tensor, cache: Dict[str, torch.Tensor] = None):
x1 = self.linear(input) # B T D
if cache is not None:
cache_layer_name = 'cache_layer_{}'.format(self.stack_layer)
if cache_layer_name not in cache:
cache[cache_layer_name] = torch.zeros(
x1.shape[0], x1.shape[-1], (self.lorder - 1) * self.lstride, 1
)
x2, cache[cache_layer_name] = self.fsmn_block(x1, cache[cache_layer_name])
else:
x2, _ = self.fsmn_block(x1, None)
x3 = self.affine(x2)
x4 = self.relu(x3)
return x4
class BasicBlock_export(nn.Module):
def __init__(
self,
model,
):
super(BasicBlock_export, self).__init__()
self.linear = model.linear
self.fsmn_block = model.fsmn_block
self.affine = model.affine
self.relu = model.relu
def forward(self, input: torch.Tensor, in_cache: torch.Tensor):
x = self.linear(input) # B T D
# cache_layer_name = 'cache_layer_{}'.format(self.stack_layer)
# if cache_layer_name not in in_cache:
# in_cache[cache_layer_name] = torch.zeros(x1.shape[0], x1.shape[-1], (self.lorder - 1) * self.lstride, 1)
x, out_cache = self.fsmn_block(x, in_cache)
x = self.affine(x)
x = self.relu(x)
return x, out_cache
class FsmnStack(nn.Sequential):
def __init__(self, *args):
super(FsmnStack, self).__init__(*args)
def forward(self, input: torch.Tensor, cache: Dict[str, torch.Tensor]):
x = input
for module in self._modules.values():
x = module(x, cache)
return x
"""
FSMN net for keyword spotting
input_dim: input dimension
linear_dim: fsmn input dimensionll
proj_dim: fsmn projection dimension
lorder: fsmn left order
rorder: fsmn right order
num_syn: output dimension
fsmn_layers: no. of sequential fsmn layers
"""
@tables.register("encoder_classes", "FSMN")
class FSMN(nn.Module):
def __init__(
self,
input_dim: int,
input_affine_dim: int,
fsmn_layers: int,
linear_dim: int,
proj_dim: int,
lorder: int,
rorder: int,
lstride: int,
rstride: int,
output_affine_dim: int,
output_dim: int,
use_softmax: bool = True,
):
super().__init__()
self.input_dim = input_dim
self.input_affine_dim = input_affine_dim
self.fsmn_layers = fsmn_layers
self.linear_dim = linear_dim
self.proj_dim = proj_dim
self.output_affine_dim = output_affine_dim
self.output_dim = output_dim
self.in_linear1 = AffineTransform(input_dim, input_affine_dim)
self.in_linear2 = AffineTransform(input_affine_dim, linear_dim)
self.relu = RectifiedLinear(linear_dim, linear_dim)
self.fsmn = FsmnStack(
*[
BasicBlock(linear_dim, proj_dim, lorder, rorder, lstride, rstride, i)
for i in range(fsmn_layers)
]
)
self.out_linear1 = AffineTransform(linear_dim, output_affine_dim)
self.out_linear2 = AffineTransform(output_affine_dim, output_dim)
self.use_softmax = use_softmax
if self.use_softmax:
self.softmax = nn.Softmax(dim=-1)
def fuse_modules(self):
pass
def output_size(self) -> int:
return self.output_dim
def forward(
self,
input: torch.Tensor,
cache: Dict[str, torch.Tensor] = None
) -> Tuple[torch.Tensor, Dict[str, torch.Tensor]]:
"""
Args:
input (torch.Tensor): Input tensor (B, T, D)
cache: when cache is not None, the forward is in streaming. The type of cache is a dict, egs,
{'cache_layer_1': torch.Tensor(B, T1, D)}, T1 is equal to self.lorder. It is {} for the 1st frame
"""
x1 = self.in_linear1(input)
x2 = self.in_linear2(x1)
x3 = self.relu(x2)
x4 = self.fsmn(x3, cache) # self.cache will update automatically in self.fsmn
x5 = self.out_linear1(x4)
x6 = self.out_linear2(x5)
if self.use_softmax:
x7 = self.softmax(x6)
return x7
return x6
@tables.register("encoder_classes", "FSMNExport")
class FSMNExport(nn.Module):
def __init__(
self,
model,
**kwargs,
):
super().__init__()
# self.input_dim = input_dim
# self.input_affine_dim = input_affine_dim
# self.fsmn_layers = fsmn_layers
# self.linear_dim = linear_dim
# self.proj_dim = proj_dim
# self.output_affine_dim = output_affine_dim
# self.output_dim = output_dim
#
# self.in_linear1 = AffineTransform(input_dim, input_affine_dim)
# self.in_linear2 = AffineTransform(input_affine_dim, linear_dim)
# self.relu = RectifiedLinear(linear_dim, linear_dim)
# self.fsmn = FsmnStack(*[BasicBlock(linear_dim, proj_dim, lorder, rorder, lstride, rstride, i) for i in
# range(fsmn_layers)])
# self.out_linear1 = AffineTransform(linear_dim, output_affine_dim)
# self.out_linear2 = AffineTransform(output_affine_dim, output_dim)
# self.softmax = nn.Softmax(dim=-1)
self.in_linear1 = model.in_linear1
self.in_linear2 = model.in_linear2
self.relu = model.relu
# self.fsmn = model.fsmn
self.out_linear1 = model.out_linear1
self.out_linear2 = model.out_linear2
self.softmax = model.softmax
self.fsmn = model.fsmn
for i, d in enumerate(model.fsmn):
if isinstance(d, BasicBlock):
self.fsmn[i] = BasicBlock_export(d)
def fuse_modules(self):
pass
def forward(
self,
input: torch.Tensor,
*args,
):
"""
Args:
input (torch.Tensor): Input tensor (B, T, D)
in_cache: when in_cache is not None, the forward is in streaming. The type of in_cache is a dict, egs,
{'cache_layer_1': torch.Tensor(B, T1, D)}, T1 is equal to self.lorder. It is {} for the 1st frame
"""
x = self.in_linear1(input)
x = self.in_linear2(x)
x = self.relu(x)
# x4 = self.fsmn(x3, in_cache) # self.in_cache will update automatically in self.fsmn
out_caches = list()
for i, d in enumerate(self.fsmn):
in_cache = args[i]
x, out_cache = d(x, in_cache)
out_caches.append(out_cache)
x = self.out_linear1(x)
x = self.out_linear2(x)
x = self.softmax(x)
return x, out_caches
|