forked from lucko515/speech-recognition-neural-network
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sample_models.py
187 lines (170 loc) · 7.89 KB
/
sample_models.py
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
from keras import backend as K
from keras.models import Model
from keras.layers import (BatchNormalization, Conv1D, Dense, Input,
TimeDistributed, Activation, Bidirectional, SimpleRNN, GRU, LSTM)
def simple_rnn_model(input_dim, output_dim=29):
""" Build a recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# Add recurrent layer
simp_rnn = GRU(output_dim, return_sequences=True,
implementation=2, name='rnn')(input_data)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(simp_rnn)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
def rnn_model(input_dim, units, activation, output_dim=29):
""" Build a recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# Add recurrent layer
simp_rnn = LSTM(units, activation=activation,
return_sequences=True, implementation=2, name='rnn')(input_data)
# TODO: Add batch normalization
bn_rnn = BatchNormalization(name='bn_rnn_1d')(simp_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bn_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
def cnn_rnn_model(input_dim, filters, kernel_size, conv_stride,
conv_border_mode, units, output_dim=29):
""" Build a recurrent + convolutional network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# Add convolutional layer
conv_1d = Conv1D(filters, kernel_size,
strides=conv_stride,
padding=conv_border_mode,
activation='relu',
name='conv1d')(input_data)
# Add batch normalization
bn_cnn = BatchNormalization(name='bn_conv_1d')(conv_1d)
# Add a recurrent layer
simp_rnn = GRU(units, activation='relu',
return_sequences=True, implementation=2, name='rnn')(bn_cnn)
# TODO: Add batch normalization
bn_rnn = BatchNormalization(name='bn_rnn_1d')(simp_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bn_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: cnn_output_length(
x, kernel_size, conv_border_mode, conv_stride)
print(model.summary())
return model
def cnn_output_length(input_length, filter_size, border_mode, stride,
dilation=1):
""" Compute the length of the output sequence after 1D convolution along
time. Note that this function is in line with the function used in
Convolution1D class from Keras.
Params:
input_length (int): Length of the input sequence.
filter_size (int): Width of the convolution kernel.
border_mode (str): Only support `same` or `valid`.
stride (int): Stride size used in 1D convolution.
dilation (int)
"""
if input_length is None:
return None
assert border_mode in {'same', 'valid'}
dilated_filter_size = filter_size + (filter_size - 1) * (dilation - 1)
if border_mode == 'same':
output_length = input_length
elif border_mode == 'valid':
output_length = input_length - dilated_filter_size + 1
return (output_length + stride - 1) // stride
def deep_rnn_model(input_dim, units, recur_layers, output_dim=29):
""" Build a deep recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# TODO: Add recurrent layers, each with batch normalization
if recur_layers == 1:
layer = LSTM(units, return_sequences=True, activation='relu')(input_data)
layer = BatchNormalization(name='bt_rnn_1')(layer)
else:
layer = LSTM(units, return_sequences=True, activation='relu')(input_data)
layer = BatchNormalization(name='bt_rnn_1')(layer)
for i in range(recur_layers - 2):
layer = LSTM(units, return_sequences=True, activation='relu')(layer)
layer = BatchNormalization(name='bt_rnn_{}'.format(2+i))(layer)
layer = LSTM(units, return_sequences=True, activation='relu')(layer)
layer = BatchNormalization(name='bt_rnn_last_rnn')(layer)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(layer)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
def bidirectional_rnn_model(input_dim, units, output_dim=29):
""" Build a bidirectional recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# TODO: Add bidirectional recurrent layer
bidir_rnn = Bidirectional(LSTM(units, return_sequences=True, activation='relu'), merge_mode='concat')(input_data)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bidir_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
def final_model(input_dim, filters, kernel_size, conv_stride,
conv_border_mode, units, output_dim=29, dropout_rate=0.5, number_of_layers=2,
cell=GRU, activation='tanh'):
""" Build a deep network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# TODO: Specify the layers in your network
conv_1d = Conv1D(filters, kernel_size,
strides=conv_stride,
padding=conv_border_mode,
activation='relu',
name='layer_1_conv',
dilation_rate=1)(input_data)
conv_bn = BatchNormalization(name='conv_batch_norm')(conv_1d)
if number_of_layers == 1:
layer = cell(units, activation=activation,
return_sequences=True, implementation=2, name='rnn_1', dropout=dropout_rate)(conv_bn)
layer = BatchNormalization(name='bt_rnn_1')(layer)
else:
layer = cell(units, activation=activation,
return_sequences=True, implementation=2, name='rnn_1', dropout=dropout_rate)(conv_bn)
layer = BatchNormalization(name='bt_rnn_1')(layer)
for i in range(number_of_layers - 2):
layer = cell(units, activation=activation,
return_sequences=True, implementation=2, name='rnn_{}'.format(i+2), dropout=dropout_rate)(layer)
layer = BatchNormalization(name='bt_rnn_{}'.format(i+2))(layer)
layer = cell(units, activation=activation,
return_sequences=True, implementation=2, name='final_layer_of_rnn')(layer)
layer = BatchNormalization(name='bt_rnn_final')(layer)
time_dense = TimeDistributed(Dense(output_dim))(layer)
# TODO: Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
# TODO: Specify model.output_length
model.output_length = lambda x: cnn_output_length(
x, kernel_size, conv_border_mode, conv_stride)
print(model.summary())
return model