| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # Copyright (c) Microsoft Corporation. |
| 4 | # Licensed under the MIT license. |
| 5 | # Code copied from Microsoft/MoE by Jacob Platin (jacobplatin@microsoft.com) |
| 6 | # but implemented by the Phi-Speech team |
| 7 | #!/usr/bin/env python3 |
| 8 | import math |
| 9 | |
| 10 | import torch |
| 11 | import torch.nn.functional as F |
| 12 | from torch import Tensor, nn |
| 13 | |
| 14 | |
| 15 | class BlockBase(nn.Module): |
| 16 | """Block abstract module.""" |
| 17 | |
| 18 | def __init__(self, input_size: int, output_size: int) -> None: |
| 19 | super().__init__() |
| 20 | self.input_size = input_size |
| 21 | self.output_size = output_size |
| 22 | |
| 23 | |
| 24 | def get_activation(name: str = "relu") -> torch.nn.Module: |
| 25 | """Select an activation function by name. |
| 26 | |
| 27 | Args: |
| 28 | name: str |
| 29 | activation function name, |
| 30 | one of ["relu", "gelu", "swish", "sigmoid"], |
| 31 | default "relu". |
| 32 | |
| 33 | """ |
| 34 | name = name.lower() |
| 35 | if name == "relu": |
| 36 | return nn.ReLU(inplace=True) |
| 37 | if name == "gelu": |
| 38 | return nn.GELU() |
| 39 | if name == "swish": |
| 40 | return nn.SiLU() |
| 41 | if name == "sigmoid": |
| 42 | return nn.Sigmoid() |
| 43 | if name == "identity": |
| 44 | return nn.Identity() |
| 45 | |
| 46 | raise NotImplementedError(name) |
| 47 | |
| 48 | |
| 49 | def adaptive_enc_mask( |
| 50 | x_len: int, chunk_start_idx: list[int], left_window: int = 0, right_window: int = 0 |
| 51 | ) -> torch.Tensor: |
| 52 | """The function is very important for Transformer Transducer Streaming mode |
| 53 | Args: |
| 54 | x_len: sequence length |
| 55 | chunk_start_idx: first idx of each chunk, such as [0,18,36,48]. |
| 56 | It also supports adaptive chunk size [0,10,15,45] |
| 57 | left_window: how many left chunks can be seen |
| 58 | right_window: how many right chunks can be seen. It is used for |
| 59 | chunk overlap model. |
| 60 | |
| 61 | Returns: |
| 62 | mask (torch.Tensor): a mask tensor for streaming model |
| 63 | Torch 1.0.1 |
| 64 | tensor([[1., 1., 0., 0.], |
| 65 | [0., 1., 1., 0.], |
| 66 | [0., 0., 1., 1.]]) |
| 67 | Torch 1.4.1 |
| 68 | tensor([[True., True., False., False.], |
| 69 | [False., True., True., False.], |
| 70 | [False., False., True., True.]]) |
| 71 | |
| 72 | """ |
| 73 | chunk_start_idx = torch.Tensor( |
| 74 | chunk_start_idx |
| 75 | ).long() # first idx of each chunk, such as [0,18,36,48]. |
| 76 | start_pad = torch.nn.functional.pad( |
| 77 | chunk_start_idx, (1, 0) |
| 78 | ) # append 0 to the beginning, so it becomes [0, 0, 18, 36, 48] |
| 79 | end_pad = torch.nn.functional.pad( |
| 80 | chunk_start_idx, (0, 1), value=x_len |
| 81 | ) # append x_len to the end, so it becomes [0,18,36,48, x_len] |
| 82 | seq_range = torch.arange(0, x_len).unsqueeze(-1) # seq_range size: [x_len, 1] |
| 83 | idx = ((seq_range < end_pad) & (seq_range >= start_pad)).nonzero()[ |
| 84 | :, 1 |
| 85 | ] # idx size: [x_len] |
| 86 | # boundary = end_pad[idx] # boundary size: [x_len] |
| 87 | seq_range_expand = ( |
| 88 | torch.arange(0, x_len).unsqueeze(0).expand(x_len, -1) |
| 89 | ) # seq_range_expand size [x_len, x_len] |
| 90 | idx_left = idx - left_window |
| 91 | idx_left[idx_left < 0] = 0 |
| 92 | boundary_left = start_pad[idx_left] |
| 93 | mask_left = seq_range_expand >= boundary_left.unsqueeze(-1) |
| 94 | idx_right = idx + right_window |
| 95 | idx_right[idx_right > len(chunk_start_idx)] = len(chunk_start_idx) |
| 96 | boundary_right = end_pad[idx_right] |
| 97 | mask_right = seq_range_expand < boundary_right.unsqueeze(-1) |
| 98 | return mask_left & mask_right |
| 99 | |
| 100 | |
| 101 | class GLU(nn.Module): |
| 102 | """Implement Gated Linear Unit (GLU) module.""" |
| 103 | |
| 104 | def __init__(self, dim: int = -1, act_name: str = "sigmoid") -> None: |
| 105 | super().__init__() |
| 106 | |
| 107 | self.dim = dim |
| 108 | self.act_fn = get_activation(act_name) |
| 109 | |
| 110 | def forward(self, x: Tensor) -> Tensor: |
| 111 | """GLU forward |
| 112 | Apply Swish function on the first half of input matrices |
| 113 | with sigmoid of the second half. |
| 114 | |
| 115 | Args: |
| 116 | x: torch.Tensor |
| 117 | Input. |
| 118 | |
| 119 | """ |
| 120 | half_x, gate = x.chunk(2, dim=self.dim) |
| 121 | return half_x * self.act_fn(gate) |
| 122 | |
| 123 | |
| 124 | # TODO: Abdel, this can be improved using GLU module |
| 125 | class GLUPointWiseConv(nn.Module): |
| 126 | """GLUPointWiseConv module |
| 127 | used for conformer architecture, |
| 128 | for more details see: |
| 129 | https://arxiv.org/pdf/2005.08100v1.pdf |
| 130 | |
| 131 | Args: |
| 132 | input_dim: int |
| 133 | input channel size. |
| 134 | output_dim: int |
| 135 | output channel size. |
| 136 | kernel_size: int |
| 137 | kernel size |
| 138 | glu_type: str, optional |
| 139 | activation function one of |
| 140 | ["sigmoid", "relu", "gelu"] |
| 141 | default "sigmoid". |
| 142 | bias_in_glu: bool, optional |
| 143 | use addtive bias in glu |
| 144 | causal: bool, optional |
| 145 | if set to True, padding is set to the half of |
| 146 | kernel size, ie, convolution can't see future frames. |
| 147 | default False. |
| 148 | |
| 149 | """ |
| 150 | |
| 151 | def __init__( |
| 152 | self, |
| 153 | input_dim: int, |
| 154 | output_dim: int, |
| 155 | kernel_size: int, |
| 156 | glu_type: str = "sigmoid", |
| 157 | bias_in_glu: bool = True, |
| 158 | causal: bool = False, |
| 159 | ) -> None: |
| 160 | super().__init__() |
| 161 | |
| 162 | self.glu_type = glu_type |
| 163 | self.output_dim = output_dim |
| 164 | self.bias_in_glu = bias_in_glu |
| 165 | if causal: |
| 166 | self.ext_pw_conv_1d = nn.Conv1d( |
| 167 | input_dim, |
| 168 | output_dim * 2, |
| 169 | kernel_size, |
| 170 | 1, |
| 171 | padding=(kernel_size - 1), |
| 172 | ) |
| 173 | else: |
| 174 | self.ext_pw_conv_1d = nn.Conv1d( |
| 175 | input_dim, |
| 176 | output_dim * 2, |
| 177 | kernel_size, |
| 178 | 1, |
| 179 | padding=(kernel_size - 1) // 2, |
| 180 | ) |
| 181 | |
| 182 | self.glu_act = get_activation(glu_type) |
| 183 | |
| 184 | if bias_in_glu: |
| 185 | self.b1 = nn.Parameter(torch.zeros(1, output_dim, 1)) |
| 186 | self.b2 = nn.Parameter(torch.zeros(1, output_dim, 1)) |
| 187 | |
| 188 | def forward(self, x: Tensor) -> Tensor: |
| 189 | """Args: |
| 190 | x: input tensor |
| 191 | |
| 192 | """ |
| 193 | # to be consistent with GLULinear, we assume the input always has the |
| 194 | # #channel (#dim) in the last dimension of the tensor, so need to |
| 195 | # switch the dimension first for 1D-Conv case |
| 196 | x = x.permute([0, 2, 1]) |
| 197 | x = self.ext_pw_conv_1d(x) |
| 198 | if self.glu_type == "bilinear": |
| 199 | if self.bias_in_glu: |
| 200 | x = (x[:, 0 : self.output_dim, :] + self.b1) * ( |
| 201 | x[:, self.output_dim : self.output_dim * 2, :] + self.b2 |
| 202 | ) |
| 203 | else: |
| 204 | x = ( |
| 205 | (x[:, 0 : self.output_dim, :]) |
| 206 | * (x[:, self.output_dim : self.output_dim * 2, :]) |
| 207 | ) |
| 208 | else: |
| 209 | if self.bias_in_glu: |
| 210 | x = (x[:, 0 : self.output_dim, :] + self.b1) * self.glu_act( |
| 211 | x[:, self.output_dim : self.output_dim * 2, :] + self.b2 |
| 212 | ) |
| 213 | else: |
| 214 | x = (x[:, 0 : self.output_dim, :]) * self.glu_act( |
| 215 | x[:, self.output_dim : self.output_dim * 2, :] |
| 216 | ) |
| 217 | |
| 218 | x = x.permute([0, 2, 1]) |
| 219 | return x |
| 220 | |
| 221 | |
| 222 | class DepthWiseSeparableConv1d(nn.Module): |
| 223 | """DepthWiseSeparableConv1d module used in ConvNet module |
| 224 | for the conformer, for more details see: |
| 225 | https://arxiv.org/pdf/2005.08100v1.pdf |
| 226 | |
| 227 | Args: |
| 228 | input_dim: int |
| 229 | input channel size. |
| 230 | depthwise_seperable_out_channel: int |
| 231 | if set different to 0, the number of |
| 232 | depthwise_seperable_out_channel will be used as a channel_out |
| 233 | of the second conv1d layer. |
| 234 | otherwise, it equals to 0, the second conv1d layer is skipped. |
| 235 | kernel_size: int |
| 236 | kernel_size |
| 237 | depthwise_multiplier: int |
| 238 | number of input_dim channels duplication. this value |
| 239 | will be used to compute the hidden channels of the Conv1D. |
| 240 | padding: int, optional |
| 241 | padding for the conv1d, |
| 242 | default: 0. |
| 243 | |
| 244 | """ |
| 245 | |
| 246 | def __init__( |
| 247 | self, |
| 248 | input_dim: int, |
| 249 | depthwise_seperable_out_channel: int, |
| 250 | kernel_size: int, |
| 251 | depthwise_multiplier: int, |
| 252 | padding: int = 0, |
| 253 | ) -> None: |
| 254 | super().__init__() |
| 255 | |
| 256 | self.dw_conv = nn.Conv1d( |
| 257 | input_dim, |
| 258 | input_dim * depthwise_multiplier, |
| 259 | kernel_size, |
| 260 | 1, |
| 261 | padding=padding, |
| 262 | groups=input_dim, |
| 263 | ) |
| 264 | |
| 265 | if depthwise_seperable_out_channel != 0: |
| 266 | self.pw_conv = nn.Conv1d( |
| 267 | input_dim * depthwise_multiplier, |
| 268 | depthwise_seperable_out_channel, |
| 269 | 1, |
| 270 | 1, |
| 271 | 0, |
| 272 | ) |
| 273 | else: |
| 274 | self.pw_conv = nn.Identity() |
| 275 | self.depthwise_seperable_out_channel = depthwise_seperable_out_channel |
| 276 | |
| 277 | def forward(self, x: Tensor) -> Tensor: |
| 278 | """Args: |
| 279 | x: input tensor |
| 280 | |
| 281 | """ |
| 282 | x = self.dw_conv(x) |
| 283 | if self.depthwise_seperable_out_channel != 0: |
| 284 | x = self.pw_conv(x) |
| 285 | return x |
| 286 | |
| 287 | |
| 288 | class ConvModule(nn.Module): |
| 289 | """ConvModule Module for the conformer block. |
| 290 | for more details see: |
| 291 | https://arxiv.org/pdf/2005.08100v1.pdf |
| 292 | |
| 293 | Args: |
| 294 | input_dim: int |
| 295 | input channel size. |
| 296 | ext_pw_out_channel: int |
| 297 | if > 0, ext_pw_out_channel is a dim channel size |
| 298 | for the last pointwise conv after swish activation. |
| 299 | depthwise_seperable_out_channel: int |
| 300 | if set different to 0, the number of |
| 301 | depthwise_seperable_out_channel |
| 302 | will be used as a channel_out of the second conv1d layer. |
| 303 | otherwise, it equal to 0, the second conv1d layer is skipped. |
| 304 | ext_pw_kernel_size: int |
| 305 | kernel size of the conv pointwise of the conformer. |
| 306 | kernel_size: int |
| 307 | kernel size. |
| 308 | depthwise_multiplier: int |
| 309 | number of input_dim channels duplication. this value |
| 310 | will be used to compute the hidden channels of the Conv1D. |
| 311 | dropout_rate: float |
| 312 | dropout rate. |
| 313 | causal: bool, optional |
| 314 | if set to True, convolution have no access |
| 315 | to future frames. default False. |
| 316 | batch_norm: bool, optional |
| 317 | if set to True, apply batchnorm before activation. |
| 318 | default False |
| 319 | chunk_se: int, optional |
| 320 | 0 for offline SE. |
| 321 | 1 for streaming SE, where mean is computed |
| 322 | by accumulated history until current chunk_se. |
| 323 | 2 for streaming SE, where mean is computed |
| 324 | by only the current chunk. |
| 325 | chunk_size: int, optional |
| 326 | chunk size for cnn. default 18 |
| 327 | activation: str, optional |
| 328 | activation function used in ConvModule, |
| 329 | default: "relu". |
| 330 | glu_type: str, optional |
| 331 | activation function used for the glu, |
| 332 | default: "sigmoid". |
| 333 | bias_in_glu: bool, optional |
| 334 | if set to True, use additive bias in the weight module |
| 335 | before GLU. |
| 336 | linear_glu_in_convm: bool, optional |
| 337 | if set to True, use GLULinear module, |
| 338 | otherwise, used GLUPointWiseConv module. |
| 339 | default to False. |
| 340 | export: bool, optional, |
| 341 | if set to True, padding is equal to 0. This is for inference, |
| 342 | or onnx export. Typically this is set by the export program or |
| 343 | the decoder program, and it isn't present in your config file. |
| 344 | default False |
| 345 | |
| 346 | """ |
| 347 | |
| 348 | def __init__( |
| 349 | self, |
| 350 | input_dim: int, |
| 351 | ext_pw_out_channel: int, |
| 352 | depthwise_seperable_out_channel: int, |
| 353 | ext_pw_kernel_size: int, |
| 354 | kernel_size: int, |
| 355 | depthwise_multiplier: int, |
| 356 | dropout_rate: float, |
| 357 | causal: bool = False, |
| 358 | batch_norm: bool = False, |
| 359 | chunk_se: int = 0, |
| 360 | chunk_size: int | list[int] = 18, |
| 361 | activation: str = "relu", |
| 362 | glu_type: str = "sigmoid", |
| 363 | bias_in_glu: bool = True, |
| 364 | linear_glu_in_convm: bool = False, |
| 365 | export: bool = False, |
| 366 | ) -> None: |
| 367 | super().__init__() |
| 368 | self.layer_norm = nn.LayerNorm(input_dim) |
| 369 | self.input_dim = input_dim |
| 370 | self.ext_pw_out_channel = ext_pw_out_channel |
| 371 | self.ext_pw_kernel_size = ext_pw_kernel_size |
| 372 | self.depthwise_seperable_out_channel = depthwise_seperable_out_channel |
| 373 | self.glu_type = glu_type |
| 374 | self.bias_in_glu = bias_in_glu |
| 375 | self.linear_glu_in_convm = linear_glu_in_convm |
| 376 | self.causal = causal |
| 377 | |
| 378 | self._add_ext_pw_layer() |
| 379 | |
| 380 | self.batch_norm = batch_norm |
| 381 | self.kernel_size = kernel_size |
| 382 | |
| 383 | if batch_norm: |
| 384 | self.bn_layer = nn.BatchNorm1d(input_dim) |
| 385 | |
| 386 | self.act = get_activation(activation) |
| 387 | self.dropout = nn.Dropout(dropout_rate) |
| 388 | self.export = export |
| 389 | |
| 390 | if causal: |
| 391 | padding = 0 if export else kernel_size - 1 |
| 392 | else: |
| 393 | padding = (kernel_size - 1) // 2 |
| 394 | |
| 395 | self.dw_sep_conv_1d = DepthWiseSeparableConv1d( |
| 396 | input_dim, |
| 397 | depthwise_seperable_out_channel, |
| 398 | kernel_size, |
| 399 | depthwise_multiplier, |
| 400 | padding=padding, |
| 401 | ) |
| 402 | |
| 403 | if depthwise_seperable_out_channel != 0: |
| 404 | if input_dim != depthwise_seperable_out_channel: |
| 405 | self.ln2 = nn.Linear(depthwise_seperable_out_channel, input_dim) |
| 406 | else: |
| 407 | if depthwise_multiplier != 1: |
| 408 | self.ln2 = nn.Linear(input_dim * depthwise_multiplier, input_dim) |
| 409 | |
| 410 | def _add_ext_pw_layer(self) -> None: |
| 411 | """This function is an extension of __init__ function |
| 412 | and dedicated to the convolution module creation |
| 413 | of the conformer. |
| 414 | """ |
| 415 | self.ln1 = self.glu = self.bn_layer = self.ext_pw_conv_1d = ( |
| 416 | nn.Identity() |
| 417 | ) # jit hacks. |
| 418 | self.squeeze_excitation = nn.Identity() # jit. |
| 419 | self.apply_ln1 = self.fix_len1 = False # jit. |
| 420 | |
| 421 | if self.ext_pw_out_channel != 0: |
| 422 | if self.causal: |
| 423 | self.ext_pw_conv_1d = nn.Conv1d( |
| 424 | self.input_dim, |
| 425 | self.ext_pw_out_channel, |
| 426 | self.ext_pw_kernel_size, |
| 427 | 1, |
| 428 | padding=(self.ext_pw_kernel_size - 1), |
| 429 | ) |
| 430 | if self.ext_pw_kernel_size > 1: |
| 431 | self.fix_len1 = True |
| 432 | else: |
| 433 | self.fix_len1 = False |
| 434 | else: |
| 435 | self.ext_pw_conv_1d = nn.Conv1d( |
| 436 | self.input_dim, |
| 437 | self.ext_pw_out_channel, |
| 438 | self.ext_pw_kernel_size, |
| 439 | 1, |
| 440 | padding=(self.ext_pw_kernel_size - 1) // 2, |
| 441 | ) |
| 442 | self.fix_len1 = False |
| 443 | |
| 444 | if self.linear_glu_in_convm: |
| 445 | self.glu = GLULinear( |
| 446 | self.input_dim, |
| 447 | self.ext_pw_out_channel, |
| 448 | self.glu_type, |
| 449 | self.bias_in_glu, |
| 450 | ) |
| 451 | else: |
| 452 | self.glu = GLUPointWiseConv( |
| 453 | self.input_dim, |
| 454 | self.ext_pw_out_channel, |
| 455 | self.ext_pw_kernel_size, |
| 456 | self.glu_type, |
| 457 | self.bias_in_glu, |
| 458 | self.causal, |
| 459 | ) |
| 460 | |
| 461 | if self.input_dim != self.ext_pw_out_channel: |
| 462 | self.apply_ln1 = True |
| 463 | self.ln1 = nn.Linear(self.ext_pw_out_channel, self.input_dim) |
| 464 | else: |
| 465 | self.apply_ln1 = False |
| 466 | else: |
| 467 | self.pw_conv_simplify_w = torch.nn.Parameter(torch.ones(3)) |
| 468 | self.pw_conv_simplify_b = torch.nn.Parameter(torch.zeros(3)) |
| 469 | |
| 470 | def forward(self, x: Tensor) -> Tensor: |
| 471 | """ConvModule Forward. |
| 472 | |
| 473 | Args: |
| 474 | x: input tensor. |
| 475 | |
| 476 | """ |
| 477 | x = self.layer_norm(x) |
| 478 | |
| 479 | if self.ext_pw_out_channel != 0: |
| 480 | x = self.glu(x) |
| 481 | if self.causal and self.ext_pw_kernel_size > 1: |
| 482 | x = x[:, : -(self.ext_pw_kernel_size - 1), :] |
| 483 | if self.apply_ln1: |
| 484 | x = self.ln1(x) |
| 485 | else: |
| 486 | x_0 = x * self.pw_conv_simplify_w[0] + self.pw_conv_simplify_b[0] |
| 487 | x_1 = x * self.pw_conv_simplify_w[1] + self.pw_conv_simplify_b[1] |
| 488 | x = x_0 + x_1 |
| 489 | |
| 490 | x = x.permute([0, 2, 1]) |
| 491 | |
| 492 | x = self.dw_sep_conv_1d(x) |
| 493 | if self.causal and self.kernel_size > 1: |
| 494 | x = x[:, :, : -(self.kernel_size - 1)] |
| 495 | if hasattr(self, "ln2"): |
| 496 | x = x.permute([0, 2, 1]) |
| 497 | x = self.ln2(x) |
| 498 | x = x.permute([0, 2, 1]) |
| 499 | if self.batch_norm: |
| 500 | x = self.bn_layer(x) |
| 501 | x = self.act(x) |
| 502 | |
| 503 | if self.ext_pw_out_channel != 0: |
| 504 | x = self.ext_pw_conv_1d(x) |
| 505 | if self.fix_len1: |
| 506 | x = x[:, :, : -(self.ext_pw_kernel_size - 1)] |
| 507 | |
| 508 | if self.apply_ln1: |
| 509 | x = x.permute([0, 2, 1]) |
| 510 | x = self.ln1(x) |
| 511 | x = x.permute([0, 2, 1]) |
| 512 | |
| 513 | x = x.permute([0, 2, 1]) |
| 514 | else: |
| 515 | x = x.unsqueeze(1).permute([0, 1, 3, 2]) |
| 516 | x = x * self.pw_conv_simplify_w[2] + self.pw_conv_simplify_b[2] |
| 517 | x = x.squeeze(1) |
| 518 | |
| 519 | x = self.dropout(x) |
| 520 | return x |
| 521 | |
| 522 | |
| 523 | class GLULinear(nn.Module): |
| 524 | """Linear + GLU module. |
| 525 | |
| 526 | Args: |
| 527 | input_dim: int |
| 528 | input size |
| 529 | output_dim: int |
| 530 | output size. |
| 531 | glu_type: |
| 532 | activation function name used in glu module. |
| 533 | default "sigmoid" (swish function). |
| 534 | bias_in_glu: bool, optional |
| 535 | If True, the addtive bias is added. Default False. |
| 536 | |
| 537 | """ |
| 538 | |
| 539 | def __init__( |
| 540 | self, |
| 541 | input_dim: int, |
| 542 | output_dim: int, |
| 543 | glu_type: str = "sigmoid", |
| 544 | bias_in_glu: bool = True, |
| 545 | ) -> None: |
| 546 | super().__init__() |
| 547 | self.linear = nn.Linear(input_dim, output_dim * 2, bias_in_glu) |
| 548 | self.glu_act = GLU(-1, glu_type) |
| 549 | |
| 550 | def forward(self, x: Tensor) -> Tensor: |
| 551 | """GLULinear forward. |
| 552 | |
| 553 | Args: |
| 554 | x: input tensor. |
| 555 | |
| 556 | """ |
| 557 | x = self.linear(x) |
| 558 | return self.glu_act(x) |
| 559 | |
| 560 | |
| 561 | class FeedForward(nn.Module): |
| 562 | """FeedForward Module. |
| 563 | For more details see Conformer paper: |
| 564 | https://arxiv.org/pdf/2005.08100.pdf |
| 565 | |
| 566 | Args: |
| 567 | d_model: int |
| 568 | input size. |
| 569 | d_inner: int |
| 570 | output size. |
| 571 | dropout_rate: float, |
| 572 | dropout rate. |
| 573 | activation: str, |
| 574 | activation function name, |
| 575 | one of ["relu", "swish", "sigmoid"], |
| 576 | sigmoid activation is only used with "glu_in_fnn=True", |
| 577 | default "sigmoid". |
| 578 | bias_in_glu: bool, optional |
| 579 | |
| 580 | """ |
| 581 | |
| 582 | def __init__( |
| 583 | self, |
| 584 | d_model: int, |
| 585 | d_inner: int, |
| 586 | dropout_rate: float, |
| 587 | activation: str = "sigmoid", |
| 588 | bias_in_glu: bool = True, |
| 589 | ) -> None: |
| 590 | super().__init__() |
| 591 | self.d_model = d_model |
| 592 | self.d_inner = d_inner |
| 593 | |
| 594 | self.layer_norm = nn.LayerNorm(d_model) |
| 595 | module = GLULinear(d_model, d_inner, activation, bias_in_glu) |
| 596 | self.net = nn.Sequential( |
| 597 | module, |
| 598 | nn.Dropout(dropout_rate), |
| 599 | nn.Linear(d_inner, d_model), |
| 600 | nn.Dropout(dropout_rate), |
| 601 | ) |
| 602 | |
| 603 | def forward(self, x: Tensor) -> Tensor: |
| 604 | """FeedForward forward function. |
| 605 | |
| 606 | Args: |
| 607 | x: input tensor. |
| 608 | |
| 609 | """ |
| 610 | out = self.net(self.layer_norm(x)) |
| 611 | |
| 612 | return out |
| 613 | |
| 614 | |
| 615 | #### positional encoding starts here |
| 616 | def _pre_hook( |
| 617 | state_dict: dict, |
| 618 | prefix: str, |
| 619 | local_metadata: dict, |
| 620 | strict: bool, |
| 621 | missing_keys: list[str], |
| 622 | unexpected_keys: list[str], |
| 623 | error_msgs: list[str], |
| 624 | ) -> None: |
| 625 | """Perform pre-hook in load_state_dict for backward compatibility. |
| 626 | |
| 627 | Note: |
| 628 | We saved self.pe until v.0.5.2 but we have omitted it later. |
| 629 | Therefore, we remove the item "pe" from `state_dict` for backward |
| 630 | compatibility. |
| 631 | |
| 632 | """ |
| 633 | k = prefix + "pe" |
| 634 | if k in state_dict: |
| 635 | state_dict.pop(k) |
| 636 | |
| 637 | |
| 638 | class T5RelativeAttentionLogitBias(nn.Module): |
| 639 | """This module implements the relative position bias described in Section |
| 640 | 2.1 of the T5 paper: https://arxiv.org/pdf/1910.10683.pdf |
| 641 | |
| 642 | The Huggingface implementation is used as a reference |
| 643 | https://github.com/huggingface/transformers/blob/v4.30.0/src/ |
| 644 | transformers/models/t5/modeling_t5.py#L435 |
| 645 | |
| 646 | Modifies attention as Q*K^T + B, where B is a learned scalar bias based |
| 647 | on relative position of the query and key. It is HxNxN, where H is the |
| 648 | number of heads, N is the sequence length. |
| 649 | |
| 650 | I've made these modifications to the original T5 bias: |
| 651 | - Skipping of the bucketing step. Original T5 bias converted rel |
| 652 | position distances into logarithmically increasing buckets. This is |
| 653 | supposed to help with length generalization. |
| 654 | - I just directly use rel position index as bias values, as we don't |
| 655 | need length generalization (40s max is good enough for ASR encoder), |
| 656 | and it keeps ONNX export simple. |
| 657 | - I've also extended it so that biases can be asymmetric, the default |
| 658 | implementation treats L->R and R->L the same. Asymmetric was found to |
| 659 | yield better results in my experiments. |
| 660 | |
| 661 | Args: |
| 662 | num_heads: int |
| 663 | Number of attention heads |
| 664 | num_buckets: int |
| 665 | Number of buckets to use for relative attention bias. This is the |
| 666 | size of the learnable bias parameter. Bucketing is not yet |
| 667 | supported, so this defaults to -1 which means no bucketing is |
| 668 | used (max_distance determines size of bias param). |
| 669 | max_distance: int |
| 670 | Maximum distance to use for relative attention bias. With |
| 671 | num_buckets=-1, this directly controls the max size of the bias |
| 672 | parameter. When num_buckets > 0 is supported, this will control |
| 673 | the maximum distance for logarithmic bucketing after which all |
| 674 | positions are in the same bucket. |
| 675 | symmetric: bool |
| 676 | Whether to use symmetric or asymmetric biases. symmetric=False uses |
| 677 | 2x number of bias params to distinguish L->R from R->L. This was |
| 678 | found to be better for the encoder. |
| 679 | |
| 680 | """ |
| 681 | |
| 682 | def __init__( |
| 683 | self, |
| 684 | num_heads: int, |
| 685 | num_buckets: int = -1, |
| 686 | max_distance: int = 1000, |
| 687 | symmetric: bool = False, |
| 688 | ) -> None: |
| 689 | super().__init__() |
| 690 | self.num_heads = num_heads |
| 691 | self.num_buckets = num_buckets |
| 692 | self.max_distance = max_distance |
| 693 | self.symmetric = symmetric |
| 694 | self._skip_bucketing = self.num_buckets < 0 |
| 695 | if self._skip_bucketing: |
| 696 | self.num_buckets = max_distance |
| 697 | else: |
| 698 | raise NotImplementedError( |
| 699 | "T5 attention bias with bucketed positions is not yet tested" |
| 700 | ) |
| 701 | if not self.symmetric: |
| 702 | self.num_buckets *= 2 |
| 703 | self.bias_values = nn.Embedding(self.num_buckets, self.num_heads) |
| 704 | |
| 705 | def forward(self, x: Tensor) -> Tensor: |
| 706 | # instantiate bias compatible with shape of x |
| 707 | maxpos = x.size(1) |
| 708 | context_position = torch.arange(maxpos, device=x.device, dtype=torch.long)[ |
| 709 | :, None |
| 710 | ] |
| 711 | memory_position = torch.arange(maxpos, device=x.device, dtype=torch.long)[ |
| 712 | None, : |
| 713 | ] |
| 714 | relative_position = memory_position - context_position |
| 715 | # clipping to a maximum distance using ops that play well with ONNX |
| 716 | # export |
| 717 | relative_position = relative_position.masked_fill( |
| 718 | relative_position < -self.max_distance, -self.max_distance |
| 719 | ) |
| 720 | relative_position = relative_position.masked_fill( |
| 721 | relative_position > self.max_distance - 1, self.max_distance - 1 |
| 722 | ) |
| 723 | |
| 724 | # mapping from relative position to index in the bias parameter |
| 725 | if self._skip_bucketing: |
| 726 | bias_idx = relative_position |
| 727 | else: |
| 728 | bias_idx = self._bucket_relative_position(relative_position) |
| 729 | if self.symmetric: |
| 730 | bias_idx = bias_idx.abs() |
| 731 | else: |
| 732 | bias_idx += self.num_buckets // 2 |
| 733 | |
| 734 | t5_rel_att_bias = self.bias_values(bias_idx) # [L, L, H] |
| 735 | t5_rel_att_bias = t5_rel_att_bias.permute(2, 0, 1).unsqueeze(0) # [1, H, L, L] |
| 736 | |
| 737 | return t5_rel_att_bias |
| 738 | |
| 739 | def _bucket_relative_position(self, relative_position: Tensor) -> Tensor: |
| 740 | # this is a placeholder (isn't tested, likely buggy) using HuggingFace |
| 741 | # implem as a reference this also needs to be extended to support |
| 742 | # asymmetric +/- ve positions |
| 743 | relative_buckets = 0 |
| 744 | if not self.causal: |
| 745 | self.num_buckets //= 2 |
| 746 | relative_buckets += (relative_position > 0).to( |
| 747 | torch.long |
| 748 | ) * self.num_buckets |
| 749 | relative_position = torch.abs(relative_position) |
| 750 | else: |
| 751 | relative_position = -torch.min( |
| 752 | relative_position, torch.zeros_like(relative_position) |
| 753 | ) |
| 754 | # now relative_position is in the range [0, inf) |
| 755 | |
| 756 | # half of the buckets are for exact increments in positions |
| 757 | max_exact = self.num_buckets // 2 |
| 758 | is_small = relative_position < max_exact |
| 759 | |
| 760 | # The other half of the buckets are for logarithmically bigger bins in |
| 761 | # positions up to max_distance |
| 762 | relative_position_if_large = max_exact + ( |
| 763 | torch.log(relative_position.float() / max_exact) |
| 764 | / math.log(self.max_distance / max_exact) |
| 765 | * (self.num_buckets - max_exact) |
| 766 | ).to(torch.long) |
| 767 | relative_position_if_large = torch.min( |
| 768 | relative_position_if_large, |
| 769 | torch.full_like(relative_position_if_large, self.num_buckets - 1), |
| 770 | ) |
| 771 | |
| 772 | relative_buckets += torch.where( |
| 773 | is_small, relative_position, relative_position_if_large |
| 774 | ) |
| 775 | return relative_buckets |
| 776 | |
| 777 | |
| 778 | class AbsolutePositionalEncoding(nn.Module): |
| 779 | """Absolute Positional encoding module. |
| 780 | This module implement Absolute sinusoidal positional encoding |
| 781 | from: https://arxiv.org/pdf/1706.03762.pdf |
| 782 | |
| 783 | Args: |
| 784 | d_model: int |
| 785 | Input embedding size. |
| 786 | dropout_rate: float |
| 787 | dropout rate |
| 788 | max_len: int, optional |
| 789 | Maximum input length sequence, Default 5000 |
| 790 | |
| 791 | """ |
| 792 | |
| 793 | def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None: |
| 794 | """Construct an PositionalEncoding object.""" |
| 795 | super().__init__() |
| 796 | self.d_model = d_model |
| 797 | self.xscale = math.sqrt(self.d_model) |
| 798 | self.dropout = torch.nn.Dropout(p=dropout_rate) |
| 799 | self.pe: torch.Tensor | None = None |
| 800 | self.extend_pe(torch.tensor(0.0).expand(1, max_len)) |
| 801 | self._register_load_state_dict_pre_hook(_pre_hook) |
| 802 | |
| 803 | def extend_pe(self, x: torch.Tensor) -> None: |
| 804 | """Reset the positional encodings. |
| 805 | |
| 806 | Args: |
| 807 | x: input tensor |
| 808 | |
| 809 | """ |
| 810 | if self.pe is not None and self.pe.size(1) >= x.size(1): |
| 811 | if self.pe.dtype != x.dtype or self.pe.device != x.device: |
| 812 | self.pe = self.pe.to(dtype=x.dtype, device=x.device) |
| 813 | return |
| 814 | pe = torch.zeros(x.size(1), self.d_model) |
| 815 | position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) |
| 816 | div_term = torch.exp( |
| 817 | torch.arange(0, self.d_model, 2, dtype=torch.float32) |
| 818 | * -(math.log(10000.0) / self.d_model) |
| 819 | ) |
| 820 | pe[:, 0::2] = torch.sin(position * div_term) |
| 821 | pe[:, 1::2] = torch.cos(position * div_term) |
| 822 | pe = pe.unsqueeze(0) |
| 823 | self.pe = pe.to(device=x.device, dtype=x.dtype) |
| 824 | |
| 825 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 826 | """Add positional encoding. |
| 827 | |
| 828 | Args: |
| 829 | x: Input tensor. shape is (batch, time, ...) |
| 830 | |
| 831 | Returns: |
| 832 | Encoded tensor. Its shape is (batch, time, ...) |
| 833 | |
| 834 | """ |
| 835 | self.extend_pe(x) |
| 836 | assert self.pe is not None |
| 837 | x = x * self.xscale + self.pe[:, : x.size(1)] |
| 838 | return self.dropout(x) |
| 839 | |
| 840 | |
| 841 | #### forward embedding layers starts here |
| 842 | class MeanVarianceNormLayer(nn.Module): |
| 843 | """Mean/variance normalization layer. |
| 844 | |
| 845 | Will subtract mean and multiply input by inverted standard deviation. |
| 846 | Typically used as a very first layer in a model. |
| 847 | |
| 848 | Args: |
| 849 | input_size: int |
| 850 | layer input size. |
| 851 | |
| 852 | """ |
| 853 | |
| 854 | def __init__(self, input_size: int) -> None: |
| 855 | super().__init__() |
| 856 | self.input_size = input_size |
| 857 | self.global_mean = nn.Parameter(torch.zeros(input_size)) |
| 858 | self.global_invstd = nn.Parameter(torch.ones(input_size)) |
| 859 | |
| 860 | def forward(self, input_: Tensor) -> Tensor: |
| 861 | """MeanVarianceNormLayer Forward. |
| 862 | |
| 863 | Args: |
| 864 | input_: input tensor. |
| 865 | |
| 866 | """ |
| 867 | return (input_ - self.global_mean) * self.global_invstd |
| 868 | |
| 869 | |
| 870 | class CausalConv1D(nn.Conv1d): |
| 871 | """A causal version of nn.Conv1d where each step would have limited access to |
| 872 | locations on its right or left |
| 873 | All arguments are the same as nn.Conv1d except padding. |
| 874 | |
| 875 | If padding is set None, then paddings are set automatically to make it a |
| 876 | causal convolution where each location would not see any steps on its right. |
| 877 | |
| 878 | If padding is set as a list (size of 2), then padding[0] would be used as |
| 879 | left padding and padding[1] as right padding. |
| 880 | It would make it possible to control the number of steps to be accessible |
| 881 | on the right and left. |
| 882 | This mode is not supported when stride > 1. padding[0]+padding[1] should |
| 883 | be equal to (kernel_size - 1). |
| 884 | """ |
| 885 | |
| 886 | def __init__( |
| 887 | self, |
| 888 | in_channels: int, |
| 889 | out_channels: int, |
| 890 | kernel_size: int, |
| 891 | stride: int = 1, |
| 892 | padding: str | int | list[int] | None = 0, |
| 893 | dilation: int = 1, |
| 894 | groups: int = 1, |
| 895 | bias: bool = True, |
| 896 | padding_mode: str = "zeros", |
| 897 | device=None, |
| 898 | dtype=None, |
| 899 | ) -> None: |
| 900 | self.cache_drop_size: int | None = None |
| 901 | self._left_padding: int |
| 902 | self._right_padding: int |
| 903 | if padding is None: |
| 904 | self._left_padding = kernel_size - 1 |
| 905 | self._right_padding = stride - 1 |
| 906 | else: |
| 907 | if stride != 1 and padding != kernel_size - 1: |
| 908 | raise ValueError("No striding allowed for non-symmetric convolutions!") |
| 909 | if isinstance(padding, int): |
| 910 | self._left_padding = padding |
| 911 | self._right_padding = padding |
| 912 | elif ( |
| 913 | isinstance(padding, list) |
| 914 | and len(padding) == 2 |
| 915 | and padding[0] + padding[1] == kernel_size - 1 |
| 916 | ): |
| 917 | self._left_padding = padding[0] |
| 918 | self._right_padding = padding[1] |
| 919 | else: |
| 920 | raise ValueError(f"Invalid padding param: {padding}!") |
| 921 | |
| 922 | self._max_cache_len = self._left_padding |
| 923 | |
| 924 | super().__init__( |
| 925 | in_channels=in_channels, |
| 926 | out_channels=out_channels, |
| 927 | kernel_size=kernel_size, |
| 928 | stride=stride, |
| 929 | padding=0, |
| 930 | dilation=dilation, |
| 931 | groups=groups, |
| 932 | bias=bias, |
| 933 | padding_mode=padding_mode, |
| 934 | device=device, |
| 935 | dtype=dtype, |
| 936 | ) |
| 937 | |
| 938 | def update_cache( |
| 939 | self, x: Tensor, cache: Tensor | None = None |
| 940 | ) -> tuple[Tensor, Tensor | None]: |
| 941 | if cache is None: |
| 942 | new_x = F.pad(x, pad=(self._left_padding, self._right_padding)) |
| 943 | next_cache = cache |
| 944 | else: |
| 945 | new_x = F.pad(x, pad=(0, self._right_padding)) |
| 946 | new_x = torch.cat([cache, new_x], dim=-1) |
| 947 | assert self.cache_drop_size is not None |
| 948 | if self.cache_drop_size > 0: |
| 949 | next_cache = new_x[:, :, : -self.cache_drop_size] |
| 950 | else: |
| 951 | next_cache = new_x |
| 952 | next_cache = next_cache[:, :, -cache.size(-1) :] |
| 953 | return new_x, next_cache |
| 954 | |
| 955 | def forward( |
| 956 | self, x: Tensor, cache: Tensor | None = None |
| 957 | ) -> Tensor | tuple[Tensor, Tensor | None]: |
| 958 | x, cache = self.update_cache(x, cache=cache) |
| 959 | x = super().forward(x) |
| 960 | if cache is None: |
| 961 | return x |
| 962 | else: |
| 963 | return x, cache |
| 964 | |
| 965 | |
| 966 | class CausalConv2D(nn.Conv2d): |
| 967 | """A causal version of nn.Conv2d where each location in the 2D matrix would |
| 968 | have no access to locations on its right or down |
| 969 | All arguments are the same as nn.Conv2d except padding which should be |
| 970 | set as None |
| 971 | """ |
| 972 | |
| 973 | def __init__( |
| 974 | self, |
| 975 | in_channels: int, |
| 976 | out_channels: int, |
| 977 | kernel_size: int, |
| 978 | stride: int = 1, |
| 979 | padding: str | int | None = 0, |
| 980 | dilation: int = 1, |
| 981 | groups: int = 1, |
| 982 | bias: bool = True, |
| 983 | padding_mode: str = "zeros", |
| 984 | device=None, |
| 985 | dtype=None, |
| 986 | ) -> None: |
| 987 | if padding is not None: |
| 988 | raise ValueError("Argument padding should be set to None for CausalConv2D.") |
| 989 | self._left_padding = kernel_size - 1 |
| 990 | self._right_padding = stride - 1 |
| 991 | |
| 992 | padding = 0 |
| 993 | super().__init__( |
| 994 | in_channels, |
| 995 | out_channels, |
| 996 | kernel_size, |
| 997 | stride, |
| 998 | padding, |
| 999 | dilation, |
| 1000 | groups, |
| 1001 | bias, |
| 1002 | padding_mode, |
| 1003 | device, |
| 1004 | dtype, |
| 1005 | ) |
| 1006 | |
| 1007 | def forward( |
| 1008 | self, |
| 1009 | x: Tensor, |
| 1010 | ) -> Tensor: |
| 1011 | x = F.pad( |
| 1012 | x, |
| 1013 | pad=(self._left_padding, self._right_padding, 0, 0), |
| 1014 | ) |
| 1015 | x = super().forward(x) |
| 1016 | return x |
| 1017 | |
| 1018 | |
| 1019 | class NemoConvSubsampling(torch.nn.Module): |
| 1020 | """Convlutional subsampling module, taken from NeMo ASR |
| 1021 | (https://github.com/NVIDIA/NeMo/blob/b367413645d5c72db3c2c96e46e95a |
| 1022 | 34501479cf/nemo/collections/asr/parts/submodules/subsampling.py) |
| 1023 | |
| 1024 | Striding Subsampling: "Speech-Transformer: A No-Recurrence |
| 1025 | Sequence-to-Sequence Model for Speech Recognition" by Linhao Dong |
| 1026 | et al. (https://ieeexplore.ieee.org/document/8462506) |
| 1027 | |
| 1028 | |
| 1029 | Compared with the EncoderConv2D (`input_layer: custom`), this is a |
| 1030 | much simplified approach, and uses no LayerNorm and far fewer Conv2Ds. |
| 1031 | Moreover, depthwise convolutions are used to reduce FLOPs, but the first |
| 1032 | layer is kept as a regular convolution so as not to degrade accuracy. |
| 1033 | |
| 1034 | `Striding` and `dw_striding` are the same except that the latter uses |
| 1035 | depthwise convolutions after the first layer, whereas the former does not. |
| 1036 | |
| 1037 | Args: |
| 1038 | subsampling_factor (int): Time reduction factor |
| 1039 | feat_in (int): size of the input features |
| 1040 | feat_out (int): size of the output features |
| 1041 | subsampling (str): The subsampling technique, choose from |
| 1042 | {"striding", "dw-striding", "striding_conv1d", |
| 1043 | "dw_striding_conv1d"} |
| 1044 | conv_channels (int): Number of channels for the convolution layers, |
| 1045 | default is 256. |
| 1046 | subsampling_conv_chunking_factor (int): Input chunking factor which |
| 1047 | can be -1 (no chunking) 1 (auto) or a power of 2. Default is 1 |
| 1048 | activation (Module): activation function, default is nn.ReLU() |
| 1049 | is_causal (bool): whether to use causal Conv1/2D, where each step will |
| 1050 | have limited access to locations on its right or left |
| 1051 | |
| 1052 | """ |
| 1053 | |
| 1054 | def __init__( |
| 1055 | self, |
| 1056 | feat_in: int, |
| 1057 | feat_out: int, |
| 1058 | subsampling_factor: int = 4, |
| 1059 | subsampling: str = "dw_striding", |
| 1060 | conv_channels: int = 256, |
| 1061 | subsampling_conv_chunking_factor: int = 1, |
| 1062 | activation: torch.nn.Module = nn.ReLU(), # noqa: B008 |
| 1063 | is_causal: bool = False, |
| 1064 | ) -> None: |
| 1065 | super().__init__() |
| 1066 | self._subsampling = subsampling |
| 1067 | self._conv_channels = conv_channels |
| 1068 | self._feat_in = feat_in |
| 1069 | self._feat_out = feat_out |
| 1070 | |
| 1071 | if subsampling_factor % 2 != 0: |
| 1072 | raise ValueError("Sampling factor should be a multiply of 2!") |
| 1073 | self._sampling_num = int(math.log(subsampling_factor, 2)) |
| 1074 | self.subsampling_factor = subsampling_factor |
| 1075 | self.is_causal = is_causal |
| 1076 | self.subsampling_causal_cond = subsampling in ( |
| 1077 | "dw_striding", |
| 1078 | "striding", |
| 1079 | "striding_conv1d", |
| 1080 | ) |
| 1081 | |
| 1082 | if ( |
| 1083 | subsampling_conv_chunking_factor != -1 |
| 1084 | and subsampling_conv_chunking_factor != 1 |
| 1085 | and subsampling_conv_chunking_factor % 2 != 0 |
| 1086 | ): |
| 1087 | raise ValueError( |
| 1088 | "subsampling_conv_chunking_factor should be -1, 1, or a power of 2" |
| 1089 | ) |
| 1090 | self.subsampling_conv_chunking_factor = subsampling_conv_chunking_factor |
| 1091 | |
| 1092 | in_channels = 1 |
| 1093 | layers = [] |
| 1094 | |
| 1095 | if subsampling == "dw_striding": |
| 1096 | self._stride = 2 |
| 1097 | self._kernel_size = 3 |
| 1098 | self._ceil_mode = False |
| 1099 | |
| 1100 | if self.is_causal: |
| 1101 | self._left_padding = self._kernel_size - 1 |
| 1102 | self._right_padding = self._stride - 1 |
| 1103 | self._max_cache_len = subsampling_factor + 1 |
| 1104 | else: |
| 1105 | self._left_padding = (self._kernel_size - 1) // 2 |
| 1106 | self._right_padding = (self._kernel_size - 1) // 2 |
| 1107 | self._max_cache_len = 0 |
| 1108 | |
| 1109 | # Layer 1 |
| 1110 | if self.is_causal: |
| 1111 | layers.append( |
| 1112 | CausalConv2D( |
| 1113 | in_channels=in_channels, |
| 1114 | out_channels=conv_channels, |
| 1115 | kernel_size=self._kernel_size, |
| 1116 | stride=self._stride, |
| 1117 | padding=None, |
| 1118 | ) |
| 1119 | ) |
| 1120 | else: |
| 1121 | layers.append( |
| 1122 | torch.nn.Conv2d( |
| 1123 | in_channels=in_channels, |
| 1124 | out_channels=conv_channels, |
| 1125 | kernel_size=self._kernel_size, |
| 1126 | stride=self._stride, |
| 1127 | padding=self._left_padding, |
| 1128 | ) |
| 1129 | ) |
| 1130 | in_channels = conv_channels |
| 1131 | layers.append(activation) |
| 1132 | |
| 1133 | for i in range(self._sampling_num - 1): |
| 1134 | if self.is_causal: |
| 1135 | layers.append( |
| 1136 | CausalConv2D( |
| 1137 | in_channels=in_channels, |
| 1138 | out_channels=in_channels, |
| 1139 | kernel_size=self._kernel_size, |
| 1140 | stride=self._stride, |
| 1141 | padding=None, |
| 1142 | groups=in_channels, |
| 1143 | ) |
| 1144 | ) |
| 1145 | else: |
| 1146 | layers.append( |
| 1147 | torch.nn.Conv2d( |
| 1148 | in_channels=in_channels, |
| 1149 | out_channels=in_channels, |
| 1150 | kernel_size=self._kernel_size, |
| 1151 | stride=self._stride, |
| 1152 | padding=self._left_padding, |
| 1153 | groups=in_channels, |
| 1154 | ) |
| 1155 | ) |
| 1156 | |
| 1157 | layers.append( |
| 1158 | torch.nn.Conv2d( |
| 1159 | in_channels=in_channels, |
| 1160 | out_channels=conv_channels, |
| 1161 | kernel_size=1, |
| 1162 | stride=1, |
| 1163 | padding=0, |
| 1164 | groups=1, |
| 1165 | ) |
| 1166 | ) |
| 1167 | layers.append(activation) |
| 1168 | in_channels = conv_channels |
| 1169 | |
| 1170 | elif subsampling == "striding": |
| 1171 | self._stride = 2 |
| 1172 | self._kernel_size = 3 |
| 1173 | self._ceil_mode = False |
| 1174 | |
| 1175 | if self.is_causal: |
| 1176 | self._left_padding = self._kernel_size - 1 |
| 1177 | self._right_padding = self._stride - 1 |
| 1178 | self._max_cache_len = subsampling_factor + 1 |
| 1179 | else: |
| 1180 | self._left_padding = (self._kernel_size - 1) // 2 |
| 1181 | self._right_padding = (self._kernel_size - 1) // 2 |
| 1182 | self._max_cache_len = 0 |
| 1183 | |
| 1184 | for i in range(self._sampling_num): |
| 1185 | if self.is_causal: |
| 1186 | layers.append( |
| 1187 | CausalConv2D( |
| 1188 | in_channels=in_channels, |
| 1189 | out_channels=conv_channels, |
| 1190 | kernel_size=self._kernel_size, |
| 1191 | stride=self._stride, |
| 1192 | padding=None, |
| 1193 | ) |
| 1194 | ) |
| 1195 | else: |
| 1196 | layers.append( |
| 1197 | torch.nn.Conv2d( |
| 1198 | in_channels=in_channels, |
| 1199 | out_channels=conv_channels, |
| 1200 | kernel_size=self._kernel_size, |
| 1201 | stride=self._stride, |
| 1202 | padding=self._left_padding, |
| 1203 | ) |
| 1204 | ) |
| 1205 | layers.append(activation) |
| 1206 | in_channels = conv_channels |
| 1207 | |
| 1208 | elif subsampling == "striding_conv1d": |
| 1209 | in_channels = feat_in |
| 1210 | |
| 1211 | self._stride = 2 |
| 1212 | self._kernel_size = 5 |
| 1213 | self._ceil_mode = False |
| 1214 | |
| 1215 | if self.is_causal: |
| 1216 | self._left_padding = self._kernel_size - 1 |
| 1217 | self._right_padding = self._stride - 1 |
| 1218 | self._max_cache_len = subsampling_factor + 1 |
| 1219 | else: |
| 1220 | self._left_padding = (self._kernel_size - 1) // 2 |
| 1221 | self._right_padding = (self._kernel_size - 1) // 2 |
| 1222 | self._max_cache_len = 0 |
| 1223 | |
| 1224 | for i in range(self._sampling_num): |
| 1225 | if self.is_causal: |
| 1226 | layers.append( |
| 1227 | CausalConv1D( |
| 1228 | in_channels=in_channels, |
| 1229 | out_channels=( |
| 1230 | feat_out |
| 1231 | if self._sampling_num == i + 1 |
| 1232 | else conv_channels |
| 1233 | ), |
| 1234 | kernel_size=self._kernel_size, |
| 1235 | stride=self._stride, |
| 1236 | padding=None, |
| 1237 | ) |
| 1238 | ) |
| 1239 | else: |
| 1240 | layers.append( |
| 1241 | torch.nn.Conv1d( |
| 1242 | in_channels=in_channels, |
| 1243 | out_channels=( |
| 1244 | feat_out |
| 1245 | if self._sampling_num == i + 1 |
| 1246 | else conv_channels |
| 1247 | ), |
| 1248 | kernel_size=self._kernel_size, |
| 1249 | stride=self._stride, |
| 1250 | padding=self._left_padding, |
| 1251 | ) |
| 1252 | ) |
| 1253 | layers.append(activation) |
| 1254 | in_channels = conv_channels |
| 1255 | |
| 1256 | elif subsampling == "dw_striding_conv1d": |
| 1257 | in_channels = feat_in |
| 1258 | |
| 1259 | self._stride = 2 |
| 1260 | self._kernel_size = 5 |
| 1261 | self._ceil_mode = False |
| 1262 | |
| 1263 | self._left_padding = (self._kernel_size - 1) // 2 |
| 1264 | self._right_padding = (self._kernel_size - 1) // 2 |
| 1265 | |
| 1266 | # Layer 1 |
| 1267 | layers.extend( |
| 1268 | [ |
| 1269 | torch.nn.Conv1d( |
| 1270 | in_channels=in_channels, |
| 1271 | out_channels=in_channels, |
| 1272 | kernel_size=self._kernel_size, |
| 1273 | stride=self._stride, |
| 1274 | padding=self._left_padding, |
| 1275 | groups=in_channels, |
| 1276 | ), |
| 1277 | torch.nn.Conv1d( |
| 1278 | in_channels=in_channels, |
| 1279 | out_channels=( |
| 1280 | feat_out if self._sampling_num == 1 else conv_channels |
| 1281 | ), |
| 1282 | kernel_size=1, |
| 1283 | stride=1, |
| 1284 | padding=0, |
| 1285 | groups=1, |
| 1286 | ), |
| 1287 | ] |
| 1288 | ) |
| 1289 | in_channels = conv_channels |
| 1290 | layers.append(activation) |
| 1291 | |
| 1292 | for i in range(self._sampling_num - 1): |
| 1293 | layers.extend( |
| 1294 | [ |
| 1295 | torch.nn.Conv1d( |
| 1296 | in_channels=in_channels, |
| 1297 | out_channels=in_channels, |
| 1298 | kernel_size=self._kernel_size, |
| 1299 | stride=self._stride, |
| 1300 | padding=self._left_padding, |
| 1301 | groups=in_channels, |
| 1302 | ), |
| 1303 | torch.nn.Conv1d( |
| 1304 | in_channels=in_channels, |
| 1305 | out_channels=( |
| 1306 | feat_out |
| 1307 | if self._sampling_num == i + 2 |
| 1308 | else conv_channels |
| 1309 | ), |
| 1310 | kernel_size=1, |
| 1311 | stride=1, |
| 1312 | padding=0, |
| 1313 | groups=1, |
| 1314 | ), |
| 1315 | ] |
| 1316 | ) |
| 1317 | layers.append(activation) |
| 1318 | in_channels = conv_channels |
| 1319 | |
| 1320 | else: |
| 1321 | raise ValueError(f"Not valid sub-sampling: {subsampling}!") |
| 1322 | |
| 1323 | if subsampling in ["dw_striding", "striding"]: |
| 1324 | out_length = calc_length_int( |
| 1325 | lengths=feat_in, |
| 1326 | all_paddings=self._left_padding + self._right_padding, |
| 1327 | kernel_size=self._kernel_size, |
| 1328 | stride=self._stride, |
| 1329 | ceil_mode=self._ceil_mode, |
| 1330 | repeat_num=self._sampling_num, |
| 1331 | ) |
| 1332 | self.out = torch.nn.Linear(conv_channels * out_length, feat_out) |
| 1333 | self.conv2d_subsampling = True |
| 1334 | elif subsampling in ["striding_conv1d", "dw_striding_conv1d"]: |
| 1335 | self.out = None |
| 1336 | self.conv2d_subsampling = False |
| 1337 | else: |
| 1338 | raise ValueError(f"Not valid sub-sampling: {subsampling}!") |
| 1339 | |
| 1340 | self.conv = torch.nn.Sequential(*layers) |
| 1341 | |
| 1342 | def get_sampling_frames(self) -> list[int]: |
| 1343 | return [1, self.subsampling_factor] |
| 1344 | |
| 1345 | def get_streaming_cache_size(self) -> list[int]: |
| 1346 | return [0, self.subsampling_factor + 1] |
| 1347 | |
| 1348 | def forward(self, x: Tensor, mask: Tensor | None) -> tuple[Tensor, Tensor | None]: |
| 1349 | """Forward method for NeMo subsampling. |
| 1350 | |
| 1351 | Args: |
| 1352 | x: input tensor |
| 1353 | mask: input mask |
| 1354 | |
| 1355 | Returns: |
| 1356 | x: Resulting tensor from subsampling (B, T // |
| 1357 | time_reduction_factor, feat_out) |
| 1358 | pad_mask: tensor of padded hidden state sequences (B, 1, T // |
| 1359 | time_reduction_factor) |
| 1360 | |
| 1361 | """ |
| 1362 | x = x.unsqueeze(1) if self.conv2d_subsampling else x.transpose(1, 2) |
| 1363 | |
| 1364 | # split inputs if chunking_factor is set |
| 1365 | if self.subsampling_conv_chunking_factor != -1 and self.conv2d_subsampling: |
| 1366 | if self.subsampling_conv_chunking_factor == 1: |
| 1367 | # if subsampling_conv_chunking_factor is 1, we split only |
| 1368 | # if needed. |
| 1369 | # avoiding a bug / feature limiting indexing of tensors |
| 1370 | # to 2**31. |
| 1371 | # see https://github.com/pytorch/pytorch/issues/80020 |
| 1372 | x_ceil = 2**31 / self._conv_channels * self._stride * self._stride |
| 1373 | need_to_split = torch.numel(x) > x_ceil |
| 1374 | else: |
| 1375 | # if subsampling_conv_chunking_factor > 1 we always split |
| 1376 | need_to_split = True |
| 1377 | |
| 1378 | if need_to_split: |
| 1379 | x, success = self.conv_split_by_batch(x) |
| 1380 | if not success: # if unable to split by batch, try by channel |
| 1381 | if self._subsampling == "dw_striding": |
| 1382 | x = self.conv_split_by_channel(x) |
| 1383 | else: |
| 1384 | x = self.conv(x) # try anyway |
| 1385 | else: |
| 1386 | x = self.conv(x) |
| 1387 | else: |
| 1388 | x = self.conv(x) |
| 1389 | |
| 1390 | # Flatten Channel and Frequency Axes |
| 1391 | if self.conv2d_subsampling: |
| 1392 | b, c, t, f = x.size() |
| 1393 | x = self.out(x.transpose(1, 2).reshape(b, t, -1)) |
| 1394 | # Transpose to Channel Last mode |
| 1395 | else: |
| 1396 | x = x.transpose(1, 2) |
| 1397 | |
| 1398 | if mask is None: |
| 1399 | return x, None |
| 1400 | |
| 1401 | max_audio_length = x.shape[1] |
| 1402 | feature_lens = mask.sum(1) |
| 1403 | padding_length = torch.ceil(feature_lens / self.subsampling_factor) |
| 1404 | if self.is_causal and self.subsampling_causal_cond: |
| 1405 | feature_lens_remainder = feature_lens % self.subsampling_factor |
| 1406 | padding_length[feature_lens_remainder != 1] += 1 |
| 1407 | pad_mask = torch.arange(0, max_audio_length, device=x.device).expand( |
| 1408 | padding_length.size(0), -1 |
| 1409 | ) < padding_length.unsqueeze(1) |
| 1410 | return x, pad_mask.unsqueeze(1) |
| 1411 | |
| 1412 | def reset_parameters(self) -> None: |
| 1413 | # initialize weights |
| 1414 | if self._subsampling == "dw_striding": |
| 1415 | with torch.no_grad(): |
| 1416 | # init conv |
| 1417 | scale = 1.0 / self._kernel_size |
| 1418 | dw_max = (self._kernel_size**2) ** -0.5 |
| 1419 | pw_max = self._conv_channels**-0.5 |
| 1420 | |
| 1421 | torch.nn.init.uniform_(self.conv[0].weight, -scale, scale) |
| 1422 | torch.nn.init.uniform_(self.conv[0].bias, -scale, scale) |
| 1423 | |
| 1424 | for idx in range(2, len(self.conv), 3): |
| 1425 | torch.nn.init.uniform_(self.conv[idx].weight, -dw_max, dw_max) |
| 1426 | torch.nn.init.uniform_(self.conv[idx].bias, -dw_max, dw_max) |
| 1427 | torch.nn.init.uniform_(self.conv[idx + 1].weight, -pw_max, pw_max) |
| 1428 | torch.nn.init.uniform_(self.conv[idx + 1].bias, -pw_max, pw_max) |
| 1429 | |
| 1430 | # init fc (80 * 64 = 5120 from https://github.com/kssteven418/ |
| 1431 | # Squeezeformer/blob/13c97d6cf92f2844d2cb3142b4c5bfa9ad1a8951/ |
| 1432 | # src/models/conformer_encoder.py#L487 |
| 1433 | fc_scale = (self._feat_out * self._feat_in / self._sampling_num) ** -0.5 |
| 1434 | torch.nn.init.uniform_(self.out.weight, -fc_scale, fc_scale) |
| 1435 | torch.nn.init.uniform_(self.out.bias, -fc_scale, fc_scale) |
| 1436 | |
| 1437 | def conv_split_by_batch(self, x: Tensor) -> tuple[Tensor, bool]: |
| 1438 | """Tries to split input by batch, run conv and concat results.""" |
| 1439 | b, _, _, _ = x.size() |
| 1440 | if b == 1: # can't split if batch size is 1 |
| 1441 | return x, False |
| 1442 | |
| 1443 | if self.subsampling_conv_chunking_factor > 1: |
| 1444 | cf = self.subsampling_conv_chunking_factor |
| 1445 | else: |
| 1446 | # avoiding a bug / feature limiting indexing of tensors to 2**31 |
| 1447 | # see https://github.com/pytorch/pytorch/issues/80020 |
| 1448 | x_ceil = 2**31 / self._conv_channels * self._stride * self._stride |
| 1449 | p = math.ceil(math.log(torch.numel(x) / x_ceil, 2)) |
| 1450 | cf = 2**p |
| 1451 | |
| 1452 | new_batch_size = b // cf |
| 1453 | if new_batch_size == 0: # input is too big |
| 1454 | return x, False |
| 1455 | |
| 1456 | return ( |
| 1457 | torch.cat( |
| 1458 | [self.conv(chunk) for chunk in torch.split(x, new_batch_size, 0)] |
| 1459 | ), |
| 1460 | True, |
| 1461 | ) |
| 1462 | |
| 1463 | def conv_split_by_channel(self, x: Tensor) -> Tensor: |
| 1464 | """For dw convs, tries to split input by time, run conv and concat |
| 1465 | results""" |
| 1466 | x = self.conv[0](x) # full conv2D |
| 1467 | x = self.conv[1](x) # activation |
| 1468 | |
| 1469 | for i in range(self._sampling_num - 1): |
| 1470 | _, c, t, _ = x.size() |
| 1471 | |
| 1472 | if self.subsampling_conv_chunking_factor > 1: |
| 1473 | cf = self.subsampling_conv_chunking_factor |
| 1474 | else: |
| 1475 | # avoiding a bug / feature limiting indexing of tensors |
| 1476 | # to 2**31 |
| 1477 | # see https://github.com/pytorch/pytorch/issues/80020 |
| 1478 | p = math.ceil(math.log(torch.numel(x) / 2**31, 2)) |
| 1479 | cf = 2**p |
| 1480 | |
| 1481 | new_c = int(c // cf) |
| 1482 | if new_c == 0: |
| 1483 | new_c = 1 |
| 1484 | |
| 1485 | new_t = int(t // cf) |
| 1486 | if new_t == 0: |
| 1487 | new_t = 1 |
| 1488 | |
| 1489 | x = self.channel_chunked_conv( |
| 1490 | self.conv[i * 3 + 2], new_c, x |
| 1491 | ) # conv2D, depthwise |
| 1492 | |
| 1493 | # splitting pointwise convs by time |
| 1494 | x = torch.cat( |
| 1495 | [self.conv[i * 3 + 3](chunk) for chunk in torch.split(x, new_t, 2)], |
| 1496 | 2, |
| 1497 | ) # conv2D, pointwise |
| 1498 | x = self.conv[i * 3 + 4](x) # activation |
| 1499 | return x |
| 1500 | |
| 1501 | def channel_chunked_conv( |
| 1502 | self, conv: torch.nn.Module, chunk_size: int, x: Tensor |
| 1503 | ) -> Tensor: |
| 1504 | """Performs channel chunked convolution.""" |
| 1505 | ind = 0 |
| 1506 | out_chunks = [] |
| 1507 | for chunk in torch.split(x, chunk_size, 1): |
| 1508 | step = chunk.size()[1] |
| 1509 | |
| 1510 | if self.is_causal: |
| 1511 | chunk = nn.functional.pad( |
| 1512 | chunk, |
| 1513 | pad=( |
| 1514 | self._kernel_size - 1, |
| 1515 | self._stride - 1, |
| 1516 | self._kernel_size - 1, |
| 1517 | self._stride - 1, |
| 1518 | ), |
| 1519 | ) |
| 1520 | ch_out = nn.functional.conv2d( |
| 1521 | chunk, |
| 1522 | conv.weight[ind : ind + step, :, :, :], |
| 1523 | bias=conv.bias[ind : ind + step], |
| 1524 | stride=self._stride, |
| 1525 | padding=0, |
| 1526 | groups=step, |
| 1527 | ) |
| 1528 | else: |
| 1529 | ch_out = nn.functional.conv2d( |
| 1530 | chunk, |
| 1531 | conv.weight[ind : ind + step, :, :, :], |
| 1532 | bias=conv.bias[ind : ind + step], |
| 1533 | stride=self._stride, |
| 1534 | padding=self._left_padding, |
| 1535 | groups=step, |
| 1536 | ) |
| 1537 | out_chunks.append(ch_out) |
| 1538 | ind += step |
| 1539 | |
| 1540 | return torch.cat(out_chunks, 1) |
| 1541 | |
| 1542 | def change_subsampling_conv_chunking_factor( |
| 1543 | self, subsampling_conv_chunking_factor: int |
| 1544 | ) -> None: |
| 1545 | if ( |
| 1546 | subsampling_conv_chunking_factor != -1 |
| 1547 | and subsampling_conv_chunking_factor != 1 |
| 1548 | and subsampling_conv_chunking_factor % 2 != 0 |
| 1549 | ): |
| 1550 | raise ValueError( |
| 1551 | "subsampling_conv_chunking_factor should be -1, 1, or a power of 2" |
| 1552 | ) |
| 1553 | self.subsampling_conv_chunking_factor = subsampling_conv_chunking_factor |
| 1554 | |
| 1555 | |
| 1556 | def calc_length_int( |
| 1557 | lengths: int, |
| 1558 | all_paddings: int, |
| 1559 | kernel_size: int, |
| 1560 | stride: int, |
| 1561 | ceil_mode: bool, |
| 1562 | repeat_num: int = 1, |
| 1563 | ) -> int: |
| 1564 | """Integer-only variant of calc_length for meta-safe shape computation. |
| 1565 | |
| 1566 | Computes the output length of a 1D convolution / pooling stack using |
| 1567 | the same formula as calc_length, but operates purely on Python numbers |
| 1568 | so it can be safely used during meta tensor initialization. |
| 1569 | """ |
| 1570 | add_pad: float = all_paddings - kernel_size |
| 1571 | one: float = 1.0 |
| 1572 | length_f: float = float(lengths) |
| 1573 | for _ in range(repeat_num): |
| 1574 | length_f = (length_f + add_pad) / stride + one |
| 1575 | length_f = math.ceil(length_f) if ceil_mode else math.floor(length_f) |
| 1576 | return int(length_f) |
| 1577 | |
| 1578 | |
| 1579 | #### multihead attention starts here |
| 1580 | class AttModule(nn.Module): |
| 1581 | """Attention abstraction module.""" |
| 1582 | |
| 1583 | def __init__(self) -> None: |
| 1584 | super().__init__() |
| 1585 | self.export_mode = False |
| 1586 | |
| 1587 | def set_export(self, mode: bool = True) -> None: |
| 1588 | """Set the export mode.""" |
| 1589 | self.export_mode = mode |
| 1590 | |
| 1591 | def forward( |
| 1592 | self, |
| 1593 | x: Tensor, |
| 1594 | memory: Tensor | None = None, |
| 1595 | pos_emb: Tensor | None = None, |
| 1596 | att_mask: Tensor | None = None, |
| 1597 | ) -> tuple[Tensor, Tensor, Tensor | None, Tensor | None]: |
| 1598 | """AttModule forward. |
| 1599 | |
| 1600 | Args: |
| 1601 | x: input tensor. |
| 1602 | memory: memory tensor. |
| 1603 | pos_emb: positional encoder embedding. |
| 1604 | att_mask: attention mask tensor. |
| 1605 | |
| 1606 | """ |
| 1607 | return x, memory, pos_emb, att_mask |
| 1608 | |
| 1609 | |
| 1610 | def masked_softmax( |
| 1611 | scores: Tensor, |
| 1612 | mask: Tensor | None, |
| 1613 | ) -> Tensor: |
| 1614 | if mask is not None: |
| 1615 | mask = mask.unsqueeze(1).eq(0) # (batch, 1, time1, time2) |
| 1616 | scores = scores.masked_fill(mask, -torch.inf) |
| 1617 | attn = torch.softmax(scores, dim=-1).masked_fill( |
| 1618 | mask, 0.0 |
| 1619 | ) # (batch, head, time1, time2) |
| 1620 | else: |
| 1621 | attn = torch.softmax(scores, dim=-1) # (batch, head, time1, time2) |
| 1622 | return attn |
| 1623 | |
| 1624 | |
| 1625 | class MultiHeadedAttention(nn.Module): |
| 1626 | """Multi-Head Attention layer with optional relative position embedding |
| 1627 | and GLU. |
| 1628 | |
| 1629 | Args: |
| 1630 | n_head: int |
| 1631 | the number of heads. |
| 1632 | n_feat: int |
| 1633 | input size features. |
| 1634 | dropout_rate: float |
| 1635 | dropout rate. |
| 1636 | attention_inner_dim: int, optional |
| 1637 | the attention dimension used in the class, |
| 1638 | it can be different from the input dimension n_feat. |
| 1639 | default: -1 (equal to n_feat). |
| 1640 | use_pt_scaled_dot_product_attention: bool, optional |
| 1641 | if set True, use pytorch scaled dot product attention in training. |
| 1642 | NOTE: this will NOT be used in ONNX decoding due to a lack of |
| 1643 | support. In that case, we use the original attention |
| 1644 | implementation, which shows no regression. |
| 1645 | default: False. |
| 1646 | n_value: int, optional |
| 1647 | if set to values other than -1, use a different dimension for |
| 1648 | value. With the default value (i.e. -1), it is backward compatible. |
| 1649 | group_size: int, optional. must divide `n_head` |
| 1650 | if group_size > 1: GQA |
| 1651 | if group_size = 1: MHA |
| 1652 | if group_size = n_head: MQA |
| 1653 | |
| 1654 | """ |
| 1655 | |
| 1656 | inv_sqrt_d_k: torch.jit.Final[float] |
| 1657 | h: torch.jit.Final[int] |
| 1658 | h_k: torch.jit.Final[int] |
| 1659 | g: torch.jit.Final[int] |
| 1660 | |
| 1661 | def __init__( |
| 1662 | self, |
| 1663 | n_head: int, |
| 1664 | n_feat: int, |
| 1665 | dropout_rate: float, |
| 1666 | attention_inner_dim: int = -1, |
| 1667 | glu_type: str = "swish", |
| 1668 | bias_in_glu: bool = True, |
| 1669 | use_pt_scaled_dot_product_attention: bool = False, |
| 1670 | n_value: int = -1, |
| 1671 | group_size: int = 1, |
| 1672 | ) -> None: |
| 1673 | super().__init__() |
| 1674 | if n_value == -1: |
| 1675 | n_value = n_feat |
| 1676 | if attention_inner_dim == -1: |
| 1677 | attention_inner_dim = n_feat |
| 1678 | assert attention_inner_dim % n_head == 0 |
| 1679 | |
| 1680 | # We assume d_v always equals d_k |
| 1681 | self.d_k = attention_inner_dim // n_head |
| 1682 | self.inv_sqrt_d_k = 1.0 / math.sqrt(self.d_k) |
| 1683 | self.h = n_head |
| 1684 | assert n_head % group_size == 0, "group_size must divide n_head" |
| 1685 | self.g = group_size |
| 1686 | self.h_k = n_head // group_size |
| 1687 | |
| 1688 | self.linear_q = nn.Linear(n_feat, attention_inner_dim) |
| 1689 | self.linear_k = nn.Linear(n_feat, attention_inner_dim // group_size) |
| 1690 | self.linear_v = nn.Linear(n_value, attention_inner_dim // group_size) |
| 1691 | self.linear_out = nn.Linear(attention_inner_dim // group_size, n_value) |
| 1692 | |
| 1693 | self.attn = torch.jit.Attribute(None, Tensor | None) |
| 1694 | self.dropout = nn.Dropout(p=dropout_rate) |
| 1695 | self.dropout_rate = dropout_rate |
| 1696 | self.use_pt_scaled_dot_product_attention = use_pt_scaled_dot_product_attention |
| 1697 | |
| 1698 | if use_pt_scaled_dot_product_attention and group_size > 1: |
| 1699 | raise ValueError("Cannot use PT Scaled Attention with GQA") |
| 1700 | |
| 1701 | # Torchscript eager quantization. Note that these functions below are |
| 1702 | # NOOPs and have very little impact on performance unless quantization |
| 1703 | # is enabled. |
| 1704 | self.quant_q = torch.ao.quantization.QuantStub() |
| 1705 | self.quant_x = torch.ao.quantization.QuantStub() |
| 1706 | self.dequant = torch.ao.quantization.DeQuantStub() |
| 1707 | self.ffunc = torch.ao.nn.quantized.FloatFunctional() |
| 1708 | |
| 1709 | def forward( |
| 1710 | self, |
| 1711 | query: Tensor, |
| 1712 | key: Tensor, |
| 1713 | value: Tensor, |
| 1714 | pos_k: Tensor | None, |
| 1715 | pos_v: Tensor | None, |
| 1716 | mask: Tensor | None, |
| 1717 | relative_attention_bias: Tensor | None = None, |
| 1718 | ) -> Tensor: |
| 1719 | """Compute 'Scaled Dot Product Attention'. |
| 1720 | |
| 1721 | Args: |
| 1722 | query: query tensor (batch, time1, size) |
| 1723 | key: key tensor (batch, time2, size) |
| 1724 | value: value tensor (batch, time1, size) |
| 1725 | pos_k: key tensor used for relative positional embedding. |
| 1726 | pos_v: value tensor used for relative positional embedding. |
| 1727 | mask: mask tensor (batch, time1, time2) |
| 1728 | relative_attention_bias: bias added to attention logits w.r.t. |
| 1729 | relative positions |
| 1730 | (1, n_head, time1, time2) |
| 1731 | |
| 1732 | """ |
| 1733 | n_batch = query.size(0) |
| 1734 | |
| 1735 | q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k) # (b, t, d) |
| 1736 | k = self.linear_k(key).view(n_batch, -1, self.h_k, self.d_k) # (b, t, d) |
| 1737 | v = self.linear_v(value).view(n_batch, -1, self.h_k, self.d_k) |
| 1738 | q = ( |
| 1739 | q.transpose(1, 2) |
| 1740 | if self.use_pt_scaled_dot_product_attention and not torch.jit.is_scripting() |
| 1741 | else q.transpose(1, 2) * self.inv_sqrt_d_k |
| 1742 | ) |
| 1743 | k = k.transpose(1, 2) # (batch, head_k, time2, d_k) |
| 1744 | v = v.transpose(1, 2) # (batch, head_k, time2, d_k) |
| 1745 | |
| 1746 | if self.use_pt_scaled_dot_product_attention and not torch.jit.is_scripting(): |
| 1747 | attn_mask = None |
| 1748 | if mask is not None: |
| 1749 | mask = mask.unsqueeze(1) |
| 1750 | if relative_attention_bias is not None: |
| 1751 | attn_mask = mask + relative_attention_bias |
| 1752 | else: |
| 1753 | attn_mask = mask |
| 1754 | if mask.dtype != q.dtype: |
| 1755 | attn_mask = attn_mask.to(q.dtype) |
| 1756 | |
| 1757 | with torch.nn.attention.sdpa_kernel( |
| 1758 | [ |
| 1759 | torch.nn.attention.SDPBackend.FLASH_ATTENTION, |
| 1760 | torch.nn.attention.SDPBackend.EFFICIENT_ATTENTION, |
| 1761 | torch.nn.attention.SDPBackend.MATH, |
| 1762 | torch.nn.attention.SDPBackend.CUDNN_ATTENTION, |
| 1763 | ] |
| 1764 | ): |
| 1765 | x = torch.nn.functional.scaled_dot_product_attention( |
| 1766 | q, |
| 1767 | k, |
| 1768 | v, |
| 1769 | attn_mask=attn_mask, |
| 1770 | dropout_p=self.dropout_rate, |
| 1771 | ) |
| 1772 | else: |
| 1773 | if self.h != self.h_k: |
| 1774 | q = q.reshape(n_batch, self.g, self.h_k, -1, self.d_k) |
| 1775 | A = torch.einsum("b g h t d, b h s d -> b h t s", q, k) |
| 1776 | else: |
| 1777 | A = torch.matmul(q, k.transpose(-2, -1)) |
| 1778 | if pos_k is not None: |
| 1779 | if self.h != self.h_k: |
| 1780 | B = torch.einsum("b g h t d, t s d -> b h t s", q, pos_k) |
| 1781 | else: |
| 1782 | reshape_q = ( |
| 1783 | q.contiguous() |
| 1784 | .view(n_batch * self.h, -1, self.d_k) |
| 1785 | .transpose(0, 1) |
| 1786 | ) # (t1,nh,dk) |
| 1787 | B = torch.matmul( |
| 1788 | reshape_q, pos_k.transpose(-2, -1) |
| 1789 | ) # pos_k: (t1,dk,t2) |
| 1790 | B = B.transpose(0, 1).view( |
| 1791 | n_batch, self.h, pos_k.size(0), pos_k.size(1) |
| 1792 | ) |
| 1793 | scores = A + B |
| 1794 | else: |
| 1795 | scores = A |
| 1796 | |
| 1797 | if relative_attention_bias is not None: |
| 1798 | scores = scores + relative_attention_bias |
| 1799 | |
| 1800 | attn = masked_softmax(scores, mask) # (batch, head, time1, time2) |
| 1801 | |
| 1802 | self.attn = attn |
| 1803 | |
| 1804 | p_attn = self.dropout(attn) |
| 1805 | x = torch.matmul(p_attn.to(v.dtype), v) # (batch, head, time1, d_k) |
| 1806 | if pos_v is not None: |
| 1807 | reshape_attn = ( |
| 1808 | p_attn.contiguous() |
| 1809 | .view(n_batch * self.h, pos_v.size(0), pos_v.size(1)) |
| 1810 | .transpose(0, 1) |
| 1811 | ) # (t1, bh, t2) |
| 1812 | |
| 1813 | attn_v = ( |
| 1814 | torch.matmul(reshape_attn, pos_v) |
| 1815 | .transpose(0, 1) |
| 1816 | .contiguous() |
| 1817 | .view(n_batch, self.h, pos_v.size(0), self.d_k) |
| 1818 | ) |
| 1819 | x = x + attn_v |
| 1820 | x = ( |
| 1821 | x.transpose(1, 2).contiguous().view(n_batch, -1, self.h_k * self.d_k) |
| 1822 | ) # (batch, time1, d_model) |
| 1823 | |
| 1824 | return self.linear_out(x) # (batch, time1, d_model) |
| 1825 | |
| 1826 | |
| 1827 | class MultiSequential(torch.nn.Sequential): |
| 1828 | """Multi-input multi-output torch.nn.Sequential.""" |
| 1829 | |
| 1830 | @torch.jit.ignore |
| 1831 | def forward(self, *args) -> tuple: |
| 1832 | """Forward method implementation.""" |
| 1833 | for m in self: |
| 1834 | args = m(*args) |
| 1835 | return args |
| 1836 | |
| 1837 | |
| 1838 | def get_offset(input_layer: str, time_reduction: int) -> int: |
| 1839 | """Get an offset. We will use the offset for determining #frames of a |
| 1840 | subsampled feature. |
| 1841 | |
| 1842 | Args: |
| 1843 | input_layer: Type of an input layer |
| 1844 | time_reduction: time reduction factor for downsampling a feature |
| 1845 | Returns: |
| 1846 | int: offset |
| 1847 | |
| 1848 | """ |
| 1849 | if input_layer in ("conv2d", "nemo_conv") and time_reduction == 4: |
| 1850 | return 3 |
| 1851 | if input_layer in ("conv2d",) and time_reduction == 6: |
| 1852 | return 1 |
| 1853 | if input_layer in ("conv2d", "nemo_conv") and time_reduction == 8: |
| 1854 | return 7 |
| 1855 | return 0 |
| 1856 | |
| 1857 | |
| 1858 | def unfold_tensor(xs_pad: Tensor, max_seq_len: int) -> Tensor: |
| 1859 | """For a given tensor with shape of (N, T, D), if sequence length T is |
| 1860 | longer than max_seq_len, this function unfold it to a |
| 1861 | (NT', max_seq_len, D) where T' is T // max_seq_len. |
| 1862 | |
| 1863 | Args: |
| 1864 | xs_pad: input tensor with shape (N, T, D) |
| 1865 | max_seq_len: maximum sequence length |
| 1866 | |
| 1867 | """ |
| 1868 | _, _, D = xs_pad.shape |
| 1869 | xs_pad = xs_pad.transpose(-1, -2) # convert to N, D, T |
| 1870 | # N x D x 1 x T => N x (D x max_seq_len) x T' |
| 1871 | xs_pad = F.unfold( |
| 1872 | xs_pad[..., None, :], |
| 1873 | kernel_size=(1, max_seq_len), |
| 1874 | stride=(1, max_seq_len), |
| 1875 | ) |
| 1876 | new_bsz, _, slen = xs_pad.shape |
| 1877 | # N x D x max_seq_len x T' |
| 1878 | xs_pad = xs_pad.view(new_bsz, -1, max_seq_len, slen) |
| 1879 | # N x T' x max_seq_len x D |
| 1880 | xs_pad = xs_pad.permute(0, 3, 2, 1).contiguous() |
| 1881 | # NT' x max_seq_len x D |
| 1882 | xs_pad = xs_pad.view(-1, max_seq_len, D) |
| 1883 | return xs_pad |