-
Notifications
You must be signed in to change notification settings - Fork 0
/
sample_models_speech_recognizer.py
303 lines (271 loc) · 12.1 KB
/
sample_models_speech_recognizer.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
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
from keras import backend as K
from keras.models import Model
from keras.layers import (BatchNormalization, Conv1D, Dense, Input,Dropout,
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 = GRU(units, activation=activation,
return_sequences=True, implementation=2, name='rnn')(input_data)
# TODO: Add batch normalization
bn_rnn = BatchNormalization(name='bn_simp_rnn')(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 = SimpleRNN(units, activation='relu',
return_sequences=True, implementation=2, name='rnn')(bn_cnn)
# TODO: Add batch normalization
bn_rnn = BatchNormalization(name='bn_rnn')(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` convolution.
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
bn_rnn = input_data
for i in range(recur_layers):
layer_name='rnn_'+str(i)
simp_rnn = GRU(units, activation='relu',
return_sequences=True, implementation=2, name=layer_name)(bn_rnn)
bn_rnn = BatchNormalization()(simp_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(units=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 deep_rnn_LSTM_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
bidir_rnn = input_data
for i in range(recur_layers):
layer_name='rnn_'+str(i)
bidir_rnn = bidir_rnn = Bidirectional(LSTM(units,
activation='relu',
return_sequences=True,
implementation=2,
name=layer_name),
merge_mode='concat')(bidir_rnn)
bidir_rnn = BatchNormalization()(bidir_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(units=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 deep_rnn_GRU_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
bidir_rnn = input_data
for i in range(recur_layers):
layer_name='rnn_'+str(i)
bidir_rnn = Bidirectional(GRU(units,
activation='relu',
return_sequences=True,
implementation=2,
name=layer_name),
merge_mode='concat')(bidir_rnn)
bidir_rnn = BatchNormalization()(bidir_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(units=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 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 = input_data
bidir_rnn = Bidirectional(LSTM(output_dim, return_sequences=True,
implementation=2, name='rnn_lstm',recurrent_dropout=0.2,dropout=0.2),
merge_mode='concat')(bidir_rnn)
# 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, recur_layers, output_dim=29):
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# TODO: Specify the layers in your network
# 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)
# Deep GRUs
nn = bn_cnn
for n in range(recur_layers):
# Add RNN layer with batch normalization
rnn = GRU(units, activation='relu', return_sequences=True, dropout = 0.3,
name='rnn_{}'.format(n))(nn)
bn_rnn = BatchNormalization(name='bn_rnn_{}'.format(n))(rnn)
nn = bn_rnn
last_layer = nn
# Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim), name='time_dense')(last_layer)
# 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 (MAKE SURE IT IS CORRECT FOR CONV1D!)
model.output_length = lambda x: cnn_output_length(
x, kernel_size, conv_border_mode, conv_stride)
print(model.summary())
return model
def final_model_1(input_dim,
# CNN parameters
#filters=200, kernel_size=11, conv_stride=2, conv_border_mode='same',
filters=350, kernel_size=11, conv_stride=1, conv_border_mode='same',
cnn_layers=3,
cnn_dropout=0.2,
cnn_activation='relu',
# RNN parameters
reccur_units=29,
recur_layers=2,
recur_implementation=2,
recurrent_dropout=0.2,
reccur_merge_mode='concat',
# Fully Connected layer parameters
fc_units=[50],
fc_dropout=0.2,
fc_activation='relu'):
""" Build a deep network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
nn=input_data
# Add convolutional layers
for i in range(cnn_layers):
layer_name='cnn_'+str(i)
nn = Conv1D(filters,
kernel_size,
strides=conv_stride,
padding=conv_border_mode,
activation=None,
name=layer_name)(nn)
nn = Activation(cnn_activation, name='act_'+layer_name)(nn)
nn = Dropout(cnn_dropout, name='drop_'+layer_name)(nn)
nn = BatchNormalization(name='bn_'+layer_name)(nn)
# TODO: Add bidirectional recurrent layers
#for i in range(recur_layers):
layer_name='bidir_rnn' #+str(i)
nn = Bidirectional(GRU(reccur_units, return_sequences=True,
implementation=recur_implementation,
name=layer_name,
dropout=0.2,
recurrent_dropout=recurrent_dropout),
merge_mode=reccur_merge_mode)(nn)
nn = BatchNormalization(name='bn_'+layer_name)(nn)
# TODO: Add a Fully Connected layers
fc_layers = len(fc_units)
for i in range(fc_layers):
layer_name='fc_'+str(i)
nn = TimeDistributed(Dense(units=fc_units[i], name=layer_name))(nn)
nn = Dropout(fc_dropout, name='drop_'+layer_name)(nn)
nn = Activation(fc_activation, name='act_'+layer_name)(nn)
nn = TimeDistributed(Dense(units=29, name='fc_out'))(nn)
# TODO: Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(nn)
# TODO: Specify the model
model = Model(inputs=input_data, outputs=y_pred)
# TODO: Specify model.output_length: select custom or Udacity version
model.output_length = lambda x: cnn_output_length(x, kernel_size, conv_border_mode, conv_stride)
print(model.summary(line_length=110))
return model