mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-09 10:02:22 +00:00
585e7b224f
* Add pruned_transducer_stateless7 for Aishell * update README.md * update comments and small fixes
88 lines
2.9 KiB
Python
88 lines
2.9 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.
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
|
|
|
|
class Decoder(nn.Module):
|
|
"""This class modifies the stateless decoder from the following paper:
|
|
|
|
RNN-transducer with stateless prediction network
|
|
https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419
|
|
|
|
It removes the recurrent connection from the decoder, i.e., the prediction
|
|
network. Different from the above paper, it adds an extra Conv1d
|
|
right after the embedding layer.
|
|
|
|
TODO: Implement https://arxiv.org/pdf/2109.07513.pdf
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
vocab_size: int,
|
|
decoder_dim: int,
|
|
blank_id: int,
|
|
context_size: int,
|
|
):
|
|
"""
|
|
Args:
|
|
vocab_size:
|
|
Number of tokens of the modeling unit including blank.
|
|
decoder_dim:
|
|
Dimension of the input embedding, and of the decoder output.
|
|
blank_id:
|
|
The ID of the blank symbol.
|
|
context_size:
|
|
Number of previous words to use to predict the next word.
|
|
1 means bigram; 2 means trigram. n means (n+1)-gram.
|
|
"""
|
|
super().__init__()
|
|
|
|
self.embedding = nn.Embedding(
|
|
num_embeddings=vocab_size,
|
|
embedding_dim=decoder_dim,
|
|
)
|
|
self.blank_id = blank_id
|
|
|
|
assert context_size == 1, context_size
|
|
self.context_size = context_size
|
|
self.vocab_size = vocab_size
|
|
|
|
def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor:
|
|
"""
|
|
Args:
|
|
y:
|
|
A 2-D tensor of shape (N, U).
|
|
need_pad:
|
|
True to left pad the input. Should be True during training.
|
|
False to not pad the input. Should be False during inference.
|
|
Returns:
|
|
Return a tensor of shape (N, U, decoder_dim).
|
|
"""
|
|
y = y.to(torch.int64)
|
|
# this stuff about clamp() is a temporary fix for a mismatch
|
|
# at utterance start, we use negative ids in beam_search.py
|
|
if torch.jit.is_tracing():
|
|
# This is for exporting to PNNX via ONNX
|
|
embedding_out = self.embedding(y)
|
|
else:
|
|
embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1)
|
|
|
|
embedding_out = F.relu(embedding_out)
|
|
return embedding_out
|