mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-08 09:32:20 +00:00
1d44da845b
* Begin to add RNN-T training for librispeech. * Copy files from conformer_ctc. Will edit it. * Use conformer/transformer model as encoder. * Begin to add training script. * Add training code. * Remove long utterances to avoid OOM when a large max_duraiton is used. * Begin to add decoding script. * Add decoding script. * Minor fixes. * Add beam search. * Use LSTM layers for the encoder. Need more tunings. * Use stateless decoder. * Minor fixes to make it ready for merge. * Fix README. * Update RESULT.md to include RNN-T Conformer. * Minor fixes. * Fix tests. * Minor fixes. * Minor fixes. * Fix tests.
102 lines
3.2 KiB
Python
102 lines
3.2 KiB
Python
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
|
|
#
|
|
# See ../../../../LICENSE for clarification regarding multiple authors
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
from typing import Optional, Tuple
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
|
|
|
|
# TODO(fangjun): Support switching between LSTM and GRU
|
|
class Decoder(nn.Module):
|
|
def __init__(
|
|
self,
|
|
vocab_size: int,
|
|
embedding_dim: int,
|
|
blank_id: int,
|
|
sos_id: int,
|
|
num_layers: int,
|
|
hidden_dim: int,
|
|
output_dim: int,
|
|
embedding_dropout: float = 0.0,
|
|
rnn_dropout: float = 0.0,
|
|
):
|
|
"""
|
|
Args:
|
|
vocab_size:
|
|
Number of tokens of the modeling unit including blank.
|
|
embedding_dim:
|
|
Dimension of the input embedding.
|
|
blank_id:
|
|
The ID of the blank symbol.
|
|
sos_id:
|
|
The ID of the SOS symbol.
|
|
num_layers:
|
|
Number of LSTM layers.
|
|
hidden_dim:
|
|
Hidden dimension of LSTM layers.
|
|
output_dim:
|
|
Output dimension of the decoder.
|
|
embedding_dropout:
|
|
Dropout rate for the embedding layer.
|
|
rnn_dropout:
|
|
Dropout for LSTM layers.
|
|
"""
|
|
super().__init__()
|
|
self.embedding = nn.Embedding(
|
|
num_embeddings=vocab_size,
|
|
embedding_dim=embedding_dim,
|
|
padding_idx=blank_id,
|
|
)
|
|
self.embedding_dropout = nn.Dropout(embedding_dropout)
|
|
# TODO(fangjun): Use layer normalized LSTM
|
|
self.rnn = nn.LSTM(
|
|
input_size=embedding_dim,
|
|
hidden_size=hidden_dim,
|
|
num_layers=num_layers,
|
|
batch_first=True,
|
|
dropout=rnn_dropout,
|
|
)
|
|
self.blank_id = blank_id
|
|
self.sos_id = sos_id
|
|
self.output_linear = nn.Linear(hidden_dim, output_dim)
|
|
|
|
def forward(
|
|
self,
|
|
y: torch.Tensor,
|
|
states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
|
) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
"""
|
|
Args:
|
|
y:
|
|
A 2-D tensor of shape (N, U) with BOS prepended.
|
|
states:
|
|
A tuple of two tensors containing the states information of
|
|
LSTM layers in this decoder.
|
|
Returns:
|
|
Return a tuple containing:
|
|
|
|
- rnn_output, a tensor of shape (N, U, C)
|
|
- (h, c), containing the state information for LSTM layers.
|
|
Both are of shape (num_layers, N, C)
|
|
"""
|
|
embeding_out = self.embedding(y)
|
|
embeding_out = self.embedding_dropout(embeding_out)
|
|
rnn_out, (h, c) = self.rnn(embeding_out, states)
|
|
out = self.output_linear(rnn_out)
|
|
|
|
return out, (h, c)
|