Spaces:
Running
Running
File size: 36,357 Bytes
864affd |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 |
from abc import ABC, abstractmethod
from typing import List, Optional, Tuple
import torch
from torchaudio.models import Emformer
__all__ = ["RNNT", "emformer_rnnt_base", "emformer_rnnt_model"]
class _TimeReduction(torch.nn.Module):
r"""Coalesces frames along time dimension into a
fewer number of frames with higher feature dimensionality.
Args:
stride (int): number of frames to merge for each output frame.
"""
def __init__(self, stride: int) -> None:
super().__init__()
self.stride = stride
def forward(self, input: torch.Tensor, lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
r"""Forward pass.
B: batch size;
T: maximum input sequence length in batch;
D: feature dimension of each input sequence frame.
Args:
input (torch.Tensor): input sequences, with shape `(B, T, D)`.
lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``input``.
Returns:
(torch.Tensor, torch.Tensor):
torch.Tensor
output sequences, with shape
`(B, T // stride, D * stride)`
torch.Tensor
output lengths, with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in output sequences.
"""
B, T, D = input.shape
num_frames = T - (T % self.stride)
input = input[:, :num_frames, :]
lengths = lengths.div(self.stride, rounding_mode="trunc")
T_max = num_frames // self.stride
output = input.reshape(B, T_max, D * self.stride)
output = output.contiguous()
return output, lengths
class _CustomLSTM(torch.nn.Module):
r"""Custom long-short-term memory (LSTM) block that applies layer normalization
to internal nodes.
Args:
input_dim (int): input dimension.
hidden_dim (int): hidden dimension.
layer_norm (bool, optional): if ``True``, enables layer normalization. (Default: ``False``)
layer_norm_epsilon (float, optional): value of epsilon to use in
layer normalization layers (Default: 1e-5)
"""
def __init__(
self,
input_dim: int,
hidden_dim: int,
layer_norm: bool = False,
layer_norm_epsilon: float = 1e-5,
) -> None:
super().__init__()
self.x2g = torch.nn.Linear(input_dim, 4 * hidden_dim, bias=(not layer_norm))
self.p2g = torch.nn.Linear(hidden_dim, 4 * hidden_dim, bias=False)
if layer_norm:
self.c_norm = torch.nn.LayerNorm(hidden_dim, eps=layer_norm_epsilon)
self.g_norm = torch.nn.LayerNorm(4 * hidden_dim, eps=layer_norm_epsilon)
else:
self.c_norm = torch.nn.Identity()
self.g_norm = torch.nn.Identity()
self.hidden_dim = hidden_dim
def forward(
self, input: torch.Tensor, state: Optional[List[torch.Tensor]]
) -> Tuple[torch.Tensor, List[torch.Tensor]]:
r"""Forward pass.
B: batch size;
T: maximum sequence length in batch;
D: feature dimension of each input sequence element.
Args:
input (torch.Tensor): with shape `(T, B, D)`.
state (List[torch.Tensor] or None): list of tensors
representing internal state generated in preceding invocation
of ``forward``.
Returns:
(torch.Tensor, List[torch.Tensor]):
torch.Tensor
output, with shape `(T, B, hidden_dim)`.
List[torch.Tensor]
list of tensors representing internal state generated
in current invocation of ``forward``.
"""
if state is None:
B = input.size(1)
h = torch.zeros(B, self.hidden_dim, device=input.device, dtype=input.dtype)
c = torch.zeros(B, self.hidden_dim, device=input.device, dtype=input.dtype)
else:
h, c = state
gated_input = self.x2g(input)
outputs = []
for gates in gated_input.unbind(0):
gates = gates + self.p2g(h)
gates = self.g_norm(gates)
input_gate, forget_gate, cell_gate, output_gate = gates.chunk(4, 1)
input_gate = input_gate.sigmoid()
forget_gate = forget_gate.sigmoid()
cell_gate = cell_gate.tanh()
output_gate = output_gate.sigmoid()
c = forget_gate * c + input_gate * cell_gate
c = self.c_norm(c)
h = output_gate * c.tanh()
outputs.append(h)
output = torch.stack(outputs, dim=0)
state = [h, c]
return output, state
class _Transcriber(ABC):
@abstractmethod
def forward(self, input: torch.Tensor, lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
pass
@abstractmethod
def infer(
self,
input: torch.Tensor,
lengths: torch.Tensor,
states: Optional[List[List[torch.Tensor]]],
) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]:
pass
class _EmformerEncoder(torch.nn.Module, _Transcriber):
r"""Emformer-based recurrent neural network transducer (RNN-T) encoder (transcription network).
Args:
input_dim (int): feature dimension of each input sequence element.
output_dim (int): feature dimension of each output sequence element.
segment_length (int): length of input segment expressed as number of frames.
right_context_length (int): length of right context expressed as number of frames.
time_reduction_input_dim (int): dimension to scale each element in input sequences to
prior to applying time reduction block.
time_reduction_stride (int): factor by which to reduce length of input sequence.
transformer_num_heads (int): number of attention heads in each Emformer layer.
transformer_ffn_dim (int): hidden layer dimension of each Emformer layer's feedforward network.
transformer_num_layers (int): number of Emformer layers to instantiate.
transformer_left_context_length (int): length of left context.
transformer_dropout (float, optional): transformer dropout probability. (Default: 0.0)
transformer_activation (str, optional): activation function to use in each Emformer layer's
feedforward network. Must be one of ("relu", "gelu", "silu"). (Default: "relu")
transformer_max_memory_size (int, optional): maximum number of memory elements to use. (Default: 0)
transformer_weight_init_scale_strategy (str, optional): per-layer weight initialization scaling
strategy. Must be one of ("depthwise", "constant", ``None``). (Default: "depthwise")
transformer_tanh_on_mem (bool, optional): if ``True``, applies tanh to memory elements. (Default: ``False``)
"""
def __init__(
self,
*,
input_dim: int,
output_dim: int,
segment_length: int,
right_context_length: int,
time_reduction_input_dim: int,
time_reduction_stride: int,
transformer_num_heads: int,
transformer_ffn_dim: int,
transformer_num_layers: int,
transformer_left_context_length: int,
transformer_dropout: float = 0.0,
transformer_activation: str = "relu",
transformer_max_memory_size: int = 0,
transformer_weight_init_scale_strategy: str = "depthwise",
transformer_tanh_on_mem: bool = False,
) -> None:
super().__init__()
self.input_linear = torch.nn.Linear(
input_dim,
time_reduction_input_dim,
bias=False,
)
self.time_reduction = _TimeReduction(time_reduction_stride)
transformer_input_dim = time_reduction_input_dim * time_reduction_stride
self.transformer = Emformer(
transformer_input_dim,
transformer_num_heads,
transformer_ffn_dim,
transformer_num_layers,
segment_length // time_reduction_stride,
dropout=transformer_dropout,
activation=transformer_activation,
left_context_length=transformer_left_context_length,
right_context_length=right_context_length // time_reduction_stride,
max_memory_size=transformer_max_memory_size,
weight_init_scale_strategy=transformer_weight_init_scale_strategy,
tanh_on_mem=transformer_tanh_on_mem,
)
self.output_linear = torch.nn.Linear(transformer_input_dim, output_dim)
self.layer_norm = torch.nn.LayerNorm(output_dim)
def forward(self, input: torch.Tensor, lengths: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
r"""Forward pass for training.
B: batch size;
T: maximum input sequence length in batch;
D: feature dimension of each input sequence frame (input_dim).
Args:
input (torch.Tensor): input frame sequences right-padded with right context, with
shape `(B, T + right context length, D)`.
lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``input``.
Returns:
(torch.Tensor, torch.Tensor):
torch.Tensor
output frame sequences, with
shape `(B, T // time_reduction_stride, output_dim)`.
torch.Tensor
output input lengths, with shape `(B,)` and i-th element representing
number of valid elements for i-th batch element in output frame sequences.
"""
input_linear_out = self.input_linear(input)
time_reduction_out, time_reduction_lengths = self.time_reduction(input_linear_out, lengths)
transformer_out, transformer_lengths = self.transformer(time_reduction_out, time_reduction_lengths)
output_linear_out = self.output_linear(transformer_out)
layer_norm_out = self.layer_norm(output_linear_out)
return layer_norm_out, transformer_lengths
@torch.jit.export
def infer(
self,
input: torch.Tensor,
lengths: torch.Tensor,
states: Optional[List[List[torch.Tensor]]],
) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]:
r"""Forward pass for inference.
B: batch size;
T: maximum input sequence segment length in batch;
D: feature dimension of each input sequence frame (input_dim).
Args:
input (torch.Tensor): input frame sequence segments right-padded with right context, with
shape `(B, T + right context length, D)`.
lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``input``.
state (List[List[torch.Tensor]] or None): list of lists of tensors
representing internal state generated in preceding invocation
of ``infer``.
Returns:
(torch.Tensor, torch.Tensor, List[List[torch.Tensor]]):
torch.Tensor
output frame sequences, with
shape `(B, T // time_reduction_stride, output_dim)`.
torch.Tensor
output input lengths, with shape `(B,)` and i-th element representing
number of valid elements for i-th batch element in output.
List[List[torch.Tensor]]
output states; list of lists of tensors
representing internal state generated in current invocation
of ``infer``.
"""
input_linear_out = self.input_linear(input)
time_reduction_out, time_reduction_lengths = self.time_reduction(input_linear_out, lengths)
(
transformer_out,
transformer_lengths,
transformer_states,
) = self.transformer.infer(time_reduction_out, time_reduction_lengths, states)
output_linear_out = self.output_linear(transformer_out)
layer_norm_out = self.layer_norm(output_linear_out)
return layer_norm_out, transformer_lengths, transformer_states
class _Predictor(torch.nn.Module):
r"""Recurrent neural network transducer (RNN-T) prediction network.
Args:
num_symbols (int): size of target token lexicon.
output_dim (int): feature dimension of each output sequence element.
symbol_embedding_dim (int): dimension of each target token embedding.
num_lstm_layers (int): number of LSTM layers to instantiate.
lstm_hidden_dim (int): output dimension of each LSTM layer.
lstm_layer_norm (bool, optional): if ``True``, enables layer normalization
for LSTM layers. (Default: ``False``)
lstm_layer_norm_epsilon (float, optional): value of epsilon to use in
LSTM layer normalization layers. (Default: 1e-5)
lstm_dropout (float, optional): LSTM dropout probability. (Default: 0.0)
"""
def __init__(
self,
num_symbols: int,
output_dim: int,
symbol_embedding_dim: int,
num_lstm_layers: int,
lstm_hidden_dim: int,
lstm_layer_norm: bool = False,
lstm_layer_norm_epsilon: float = 1e-5,
lstm_dropout: float = 0.0,
) -> None:
super().__init__()
self.embedding = torch.nn.Embedding(num_symbols, symbol_embedding_dim)
self.input_layer_norm = torch.nn.LayerNorm(symbol_embedding_dim)
self.lstm_layers = torch.nn.ModuleList(
[
_CustomLSTM(
symbol_embedding_dim if idx == 0 else lstm_hidden_dim,
lstm_hidden_dim,
layer_norm=lstm_layer_norm,
layer_norm_epsilon=lstm_layer_norm_epsilon,
)
for idx in range(num_lstm_layers)
]
)
self.dropout = torch.nn.Dropout(p=lstm_dropout)
self.linear = torch.nn.Linear(lstm_hidden_dim, output_dim)
self.output_layer_norm = torch.nn.LayerNorm(output_dim)
self.lstm_dropout = lstm_dropout
def forward(
self,
input: torch.Tensor,
lengths: torch.Tensor,
state: Optional[List[List[torch.Tensor]]] = None,
) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]:
r"""Forward pass.
B: batch size;
U: maximum sequence length in batch;
D: feature dimension of each input sequence element.
Args:
input (torch.Tensor): target sequences, with shape `(B, U)` and each element
mapping to a target symbol, i.e. in range `[0, num_symbols)`.
lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``input``.
state (List[List[torch.Tensor]] or None, optional): list of lists of tensors
representing internal state generated in preceding invocation
of ``forward``. (Default: ``None``)
Returns:
(torch.Tensor, torch.Tensor, List[List[torch.Tensor]]):
torch.Tensor
output encoding sequences, with shape `(B, U, output_dim)`
torch.Tensor
output lengths, with shape `(B,)` and i-th element representing
number of valid elements for i-th batch element in output encoding sequences.
List[List[torch.Tensor]]
output states; list of lists of tensors
representing internal state generated in current invocation of ``forward``.
"""
input_tb = input.permute(1, 0)
embedding_out = self.embedding(input_tb)
input_layer_norm_out = self.input_layer_norm(embedding_out)
lstm_out = input_layer_norm_out
state_out: List[List[torch.Tensor]] = []
for layer_idx, lstm in enumerate(self.lstm_layers):
lstm_out, lstm_state_out = lstm(lstm_out, None if state is None else state[layer_idx])
lstm_out = self.dropout(lstm_out)
state_out.append(lstm_state_out)
linear_out = self.linear(lstm_out)
output_layer_norm_out = self.output_layer_norm(linear_out)
return output_layer_norm_out.permute(1, 0, 2), lengths, state_out
class _Joiner(torch.nn.Module):
r"""Recurrent neural network transducer (RNN-T) joint network.
Args:
input_dim (int): source and target input dimension.
output_dim (int): output dimension.
activation (str, optional): activation function to use in the joiner.
Must be one of ("relu", "tanh"). (Default: "relu")
"""
def __init__(self, input_dim: int, output_dim: int, activation: str = "relu") -> None:
super().__init__()
self.linear = torch.nn.Linear(input_dim, output_dim, bias=True)
if activation == "relu":
self.activation = torch.nn.ReLU()
elif activation == "tanh":
self.activation = torch.nn.Tanh()
else:
raise ValueError(f"Unsupported activation {activation}")
def forward(
self,
source_encodings: torch.Tensor,
source_lengths: torch.Tensor,
target_encodings: torch.Tensor,
target_lengths: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
r"""Forward pass for training.
B: batch size;
T: maximum source sequence length in batch;
U: maximum target sequence length in batch;
D: dimension of each source and target sequence encoding.
Args:
source_encodings (torch.Tensor): source encoding sequences, with
shape `(B, T, D)`.
source_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
valid sequence length of i-th batch element in ``source_encodings``.
target_encodings (torch.Tensor): target encoding sequences, with shape `(B, U, D)`.
target_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
valid sequence length of i-th batch element in ``target_encodings``.
Returns:
(torch.Tensor, torch.Tensor, torch.Tensor):
torch.Tensor
joint network output, with shape `(B, T, U, output_dim)`.
torch.Tensor
output source lengths, with shape `(B,)` and i-th element representing
number of valid elements along dim 1 for i-th batch element in joint network output.
torch.Tensor
output target lengths, with shape `(B,)` and i-th element representing
number of valid elements along dim 2 for i-th batch element in joint network output.
"""
joint_encodings = source_encodings.unsqueeze(2).contiguous() + target_encodings.unsqueeze(1).contiguous()
activation_out = self.activation(joint_encodings)
output = self.linear(activation_out)
return output, source_lengths, target_lengths
class RNNT(torch.nn.Module):
r"""torchaudio.models.RNNT()
Recurrent neural network transducer (RNN-T) model.
Note:
To build the model, please use one of the factory functions.
See Also:
:class:`torchaudio.pipelines.RNNTBundle`: ASR pipeline with pre-trained models.
Args:
transcriber (torch.nn.Module): transcription network.
predictor (torch.nn.Module): prediction network.
joiner (torch.nn.Module): joint network.
"""
def __init__(self, transcriber: _Transcriber, predictor: _Predictor, joiner: _Joiner) -> None:
super().__init__()
self.transcriber = transcriber
self.predictor = predictor
self.joiner = joiner
def forward(
self,
sources: torch.Tensor,
source_lengths: torch.Tensor,
targets: torch.Tensor,
target_lengths: torch.Tensor,
predictor_state: Optional[List[List[torch.Tensor]]] = None,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]:
r"""Forward pass for training.
B: batch size;
T: maximum source sequence length in batch;
U: maximum target sequence length in batch;
D: feature dimension of each source sequence element.
Args:
sources (torch.Tensor): source frame sequences right-padded with right context, with
shape `(B, T, D)`.
source_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``sources``.
targets (torch.Tensor): target sequences, with shape `(B, U)` and each element
mapping to a target symbol.
target_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``targets``.
predictor_state (List[List[torch.Tensor]] or None, optional): list of lists of tensors
representing prediction network internal state generated in preceding invocation
of ``forward``. (Default: ``None``)
Returns:
(torch.Tensor, torch.Tensor, torch.Tensor, List[List[torch.Tensor]]):
torch.Tensor
joint network output, with shape
`(B, max output source length, max output target length, output_dim (number of target symbols))`.
torch.Tensor
output source lengths, with shape `(B,)` and i-th element representing
number of valid elements along dim 1 for i-th batch element in joint network output.
torch.Tensor
output target lengths, with shape `(B,)` and i-th element representing
number of valid elements along dim 2 for i-th batch element in joint network output.
List[List[torch.Tensor]]
output states; list of lists of tensors
representing prediction network internal state generated in current invocation
of ``forward``.
"""
source_encodings, source_lengths = self.transcriber(
input=sources,
lengths=source_lengths,
)
target_encodings, target_lengths, predictor_state = self.predictor(
input=targets,
lengths=target_lengths,
state=predictor_state,
)
output, source_lengths, target_lengths = self.joiner(
source_encodings=source_encodings,
source_lengths=source_lengths,
target_encodings=target_encodings,
target_lengths=target_lengths,
)
return (
output,
source_lengths,
target_lengths,
predictor_state,
)
@torch.jit.export
def transcribe_streaming(
self,
sources: torch.Tensor,
source_lengths: torch.Tensor,
state: Optional[List[List[torch.Tensor]]],
) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]:
r"""Applies transcription network to sources in streaming mode.
B: batch size;
T: maximum source sequence segment length in batch;
D: feature dimension of each source sequence frame.
Args:
sources (torch.Tensor): source frame sequence segments right-padded with right context, with
shape `(B, T + right context length, D)`.
source_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``sources``.
state (List[List[torch.Tensor]] or None): list of lists of tensors
representing transcription network internal state generated in preceding invocation
of ``transcribe_streaming``.
Returns:
(torch.Tensor, torch.Tensor, List[List[torch.Tensor]]):
torch.Tensor
output frame sequences, with
shape `(B, T // time_reduction_stride, output_dim)`.
torch.Tensor
output lengths, with shape `(B,)` and i-th element representing
number of valid elements for i-th batch element in output.
List[List[torch.Tensor]]
output states; list of lists of tensors
representing transcription network internal state generated in current invocation
of ``transcribe_streaming``.
"""
return self.transcriber.infer(sources, source_lengths, state)
@torch.jit.export
def transcribe(
self,
sources: torch.Tensor,
source_lengths: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor]:
r"""Applies transcription network to sources in non-streaming mode.
B: batch size;
T: maximum source sequence length in batch;
D: feature dimension of each source sequence frame.
Args:
sources (torch.Tensor): source frame sequences right-padded with right context, with
shape `(B, T + right context length, D)`.
source_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``sources``.
Returns:
(torch.Tensor, torch.Tensor):
torch.Tensor
output frame sequences, with
shape `(B, T // time_reduction_stride, output_dim)`.
torch.Tensor
output lengths, with shape `(B,)` and i-th element representing
number of valid elements for i-th batch element in output frame sequences.
"""
return self.transcriber(sources, source_lengths)
@torch.jit.export
def predict(
self,
targets: torch.Tensor,
target_lengths: torch.Tensor,
state: Optional[List[List[torch.Tensor]]],
) -> Tuple[torch.Tensor, torch.Tensor, List[List[torch.Tensor]]]:
r"""Applies prediction network to targets.
B: batch size;
U: maximum target sequence length in batch;
D: feature dimension of each target sequence frame.
Args:
targets (torch.Tensor): target sequences, with shape `(B, U)` and each element
mapping to a target symbol, i.e. in range `[0, num_symbols)`.
target_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
number of valid frames for i-th batch element in ``targets``.
state (List[List[torch.Tensor]] or None): list of lists of tensors
representing internal state generated in preceding invocation
of ``predict``.
Returns:
(torch.Tensor, torch.Tensor, List[List[torch.Tensor]]):
torch.Tensor
output frame sequences, with shape `(B, U, output_dim)`.
torch.Tensor
output lengths, with shape `(B,)` and i-th element representing
number of valid elements for i-th batch element in output.
List[List[torch.Tensor]]
output states; list of lists of tensors
representing internal state generated in current invocation of ``predict``.
"""
return self.predictor(input=targets, lengths=target_lengths, state=state)
@torch.jit.export
def join(
self,
source_encodings: torch.Tensor,
source_lengths: torch.Tensor,
target_encodings: torch.Tensor,
target_lengths: torch.Tensor,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
r"""Applies joint network to source and target encodings.
B: batch size;
T: maximum source sequence length in batch;
U: maximum target sequence length in batch;
D: dimension of each source and target sequence encoding.
Args:
source_encodings (torch.Tensor): source encoding sequences, with
shape `(B, T, D)`.
source_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
valid sequence length of i-th batch element in ``source_encodings``.
target_encodings (torch.Tensor): target encoding sequences, with shape `(B, U, D)`.
target_lengths (torch.Tensor): with shape `(B,)` and i-th element representing
valid sequence length of i-th batch element in ``target_encodings``.
Returns:
(torch.Tensor, torch.Tensor, torch.Tensor):
torch.Tensor
joint network output, with shape `(B, T, U, output_dim)`.
torch.Tensor
output source lengths, with shape `(B,)` and i-th element representing
number of valid elements along dim 1 for i-th batch element in joint network output.
torch.Tensor
output target lengths, with shape `(B,)` and i-th element representing
number of valid elements along dim 2 for i-th batch element in joint network output.
"""
output, source_lengths, target_lengths = self.joiner(
source_encodings=source_encodings,
source_lengths=source_lengths,
target_encodings=target_encodings,
target_lengths=target_lengths,
)
return output, source_lengths, target_lengths
def emformer_rnnt_model(
*,
input_dim: int,
encoding_dim: int,
num_symbols: int,
segment_length: int,
right_context_length: int,
time_reduction_input_dim: int,
time_reduction_stride: int,
transformer_num_heads: int,
transformer_ffn_dim: int,
transformer_num_layers: int,
transformer_dropout: float,
transformer_activation: str,
transformer_left_context_length: int,
transformer_max_memory_size: int,
transformer_weight_init_scale_strategy: str,
transformer_tanh_on_mem: bool,
symbol_embedding_dim: int,
num_lstm_layers: int,
lstm_layer_norm: bool,
lstm_layer_norm_epsilon: float,
lstm_dropout: float,
) -> RNNT:
r"""Builds Emformer-based :class:`~torchaudio.models.RNNT`.
Note:
For non-streaming inference, the expectation is for `transcribe` to be called on input
sequences right-concatenated with `right_context_length` frames.
For streaming inference, the expectation is for `transcribe_streaming` to be called
on input chunks comprising `segment_length` frames right-concatenated with `right_context_length`
frames.
Args:
input_dim (int): dimension of input sequence frames passed to transcription network.
encoding_dim (int): dimension of transcription- and prediction-network-generated encodings
passed to joint network.
num_symbols (int): cardinality of set of target tokens.
segment_length (int): length of input segment expressed as number of frames.
right_context_length (int): length of right context expressed as number of frames.
time_reduction_input_dim (int): dimension to scale each element in input sequences to
prior to applying time reduction block.
time_reduction_stride (int): factor by which to reduce length of input sequence.
transformer_num_heads (int): number of attention heads in each Emformer layer.
transformer_ffn_dim (int): hidden layer dimension of each Emformer layer's feedforward network.
transformer_num_layers (int): number of Emformer layers to instantiate.
transformer_left_context_length (int): length of left context considered by Emformer.
transformer_dropout (float): Emformer dropout probability.
transformer_activation (str): activation function to use in each Emformer layer's
feedforward network. Must be one of ("relu", "gelu", "silu").
transformer_max_memory_size (int): maximum number of memory elements to use.
transformer_weight_init_scale_strategy (str): per-layer weight initialization scaling
strategy. Must be one of ("depthwise", "constant", ``None``).
transformer_tanh_on_mem (bool): if ``True``, applies tanh to memory elements.
symbol_embedding_dim (int): dimension of each target token embedding.
num_lstm_layers (int): number of LSTM layers to instantiate.
lstm_layer_norm (bool): if ``True``, enables layer normalization for LSTM layers.
lstm_layer_norm_epsilon (float): value of epsilon to use in LSTM layer normalization layers.
lstm_dropout (float): LSTM dropout probability.
Returns:
RNNT:
Emformer RNN-T model.
"""
encoder = _EmformerEncoder(
input_dim=input_dim,
output_dim=encoding_dim,
segment_length=segment_length,
right_context_length=right_context_length,
time_reduction_input_dim=time_reduction_input_dim,
time_reduction_stride=time_reduction_stride,
transformer_num_heads=transformer_num_heads,
transformer_ffn_dim=transformer_ffn_dim,
transformer_num_layers=transformer_num_layers,
transformer_dropout=transformer_dropout,
transformer_activation=transformer_activation,
transformer_left_context_length=transformer_left_context_length,
transformer_max_memory_size=transformer_max_memory_size,
transformer_weight_init_scale_strategy=transformer_weight_init_scale_strategy,
transformer_tanh_on_mem=transformer_tanh_on_mem,
)
predictor = _Predictor(
num_symbols,
encoding_dim,
symbol_embedding_dim=symbol_embedding_dim,
num_lstm_layers=num_lstm_layers,
lstm_hidden_dim=symbol_embedding_dim,
lstm_layer_norm=lstm_layer_norm,
lstm_layer_norm_epsilon=lstm_layer_norm_epsilon,
lstm_dropout=lstm_dropout,
)
joiner = _Joiner(encoding_dim, num_symbols)
return RNNT(encoder, predictor, joiner)
def emformer_rnnt_base(num_symbols: int) -> RNNT:
r"""Builds basic version of Emformer-based :class:`~torchaudio.models.RNNT`.
Args:
num_symbols (int): The size of target token lexicon.
Returns:
RNNT:
Emformer RNN-T model.
"""
return emformer_rnnt_model(
input_dim=80,
encoding_dim=1024,
num_symbols=num_symbols,
segment_length=16,
right_context_length=4,
time_reduction_input_dim=128,
time_reduction_stride=4,
transformer_num_heads=8,
transformer_ffn_dim=2048,
transformer_num_layers=20,
transformer_dropout=0.1,
transformer_activation="gelu",
transformer_left_context_length=30,
transformer_max_memory_size=0,
transformer_weight_init_scale_strategy="depthwise",
transformer_tanh_on_mem=True,
symbol_embedding_dim=512,
num_lstm_layers=3,
lstm_layer_norm=True,
lstm_layer_norm_epsilon=1e-3,
lstm_dropout=0.3,
)
|