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
|
"""Main definition of the Deep speech 2 model by Baidu Research.
Following definition by Assembly AI
(https://www.assemblyai.com/blog/end-to-end-speech-recognition-pytorch/)
"""
from typing import TypedDict
import torch.nn.functional as F
from torch import nn
class HParams(TypedDict):
"""Type for the hyperparameters of the model."""
n_cnn_layers: int
n_rnn_layers: int
rnn_dim: int
n_class: int
n_feats: int
stride: int
dropout: float
learning_rate: float
batch_size: int
epochs: int
class CNNLayerNorm(nn.Module):
"""Layer normalization built for cnns input"""
def __init__(self, n_feats: int):
super().__init__()
self.layer_norm = nn.LayerNorm(n_feats)
def forward(self, data):
"""x (batch, channel, feature, time)"""
data = data.transpose(2, 3).contiguous() # (batch, channel, time, feature)
data = self.layer_norm(data)
return data.transpose(2, 3).contiguous() # (batch, channel, feature, time)
class ResidualCNN(nn.Module):
"""Residual CNN inspired by https://arxiv.org/pdf/1603.05027.pdf"""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel: int,
stride: int,
dropout: float,
n_feats: int,
):
super().__init__()
self.cnn1 = nn.Conv2d(in_channels, out_channels, kernel, stride, padding=kernel // 2)
self.cnn2 = nn.Conv2d(
out_channels,
out_channels,
kernel,
stride,
padding=kernel // 2,
)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.layer_norm1 = CNNLayerNorm(n_feats)
self.layer_norm2 = CNNLayerNorm(n_feats)
def forward(self, data):
"""x (batch, channel, feature, time)"""
residual = data # (batch, channel, feature, time)
data = self.layer_norm1(data)
data = F.gelu(data)
data = self.dropout1(data)
data = self.cnn1(data)
data = self.layer_norm2(data)
data = F.gelu(data)
data = self.dropout2(data)
data = self.cnn2(data)
data += residual
return data # (batch, channel, feature, time)
class BidirectionalGRU(nn.Module):
"""Bidirectional GRU with Layer Normalization and Dropout"""
def __init__(
self,
rnn_dim: int,
hidden_size: int,
dropout: float,
batch_first: bool,
):
super().__init__()
self.bi_gru = nn.GRU(
input_size=rnn_dim,
hidden_size=hidden_size,
num_layers=1,
batch_first=batch_first,
bidirectional=True,
)
self.layer_norm = nn.LayerNorm(rnn_dim)
self.dropout = nn.Dropout(dropout)
def forward(self, data):
"""data (batch, time, feature)"""
data = self.layer_norm(data)
data = F.gelu(data)
data = self.dropout(data)
data, _ = self.bi_gru(data)
return data
class SpeechRecognitionModel(nn.Module):
"""Speech Recognition Model Inspired by DeepSpeech 2"""
def __init__(
self,
n_cnn_layers: int,
n_rnn_layers: int,
rnn_dim: int,
n_class: int,
n_feats: int,
stride: int = 2,
dropout: float = 0.1,
):
super().__init__()
n_feats //= 2
self.cnn = nn.Conv2d(1, 32, 3, stride=stride, padding=3 // 2)
# n residual cnn layers with filter size of 32
self.rescnn_layers = nn.Sequential(
*[
ResidualCNN(32, 32, kernel=3, stride=1, dropout=dropout, n_feats=n_feats)
for _ in range(n_cnn_layers)
]
)
self.fully_connected = nn.Linear(n_feats * 32, rnn_dim)
self.birnn_layers = nn.Sequential(
*[
BidirectionalGRU(
rnn_dim=rnn_dim if i == 0 else rnn_dim * 2,
hidden_size=rnn_dim,
dropout=dropout,
batch_first=i == 0,
)
for i in range(n_rnn_layers)
]
)
self.classifier = nn.Sequential(
nn.Linear(rnn_dim * 2, rnn_dim), # birnn returns rnn_dim*2
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(rnn_dim, n_class),
)
def forward(self, data):
"""data (batch, channel, feature, time)"""
data = self.cnn(data)
data = self.rescnn_layers(data)
sizes = data.size()
data = data.view(sizes[0], sizes[1] * sizes[2], sizes[3]) # (batch, feature, time)
data = data.transpose(1, 2) # (batch, time, feature)
data = self.fully_connected(data)
data = self.birnn_layers(data)
data = self.classifier(data)
return data
|
