-
Notifications
You must be signed in to change notification settings - Fork 12
/
WaveNetClassifier.py
188 lines (171 loc) · 7.63 KB
/
WaveNetClassifier.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
import numpy as np
import pandas as pd
import sys
import tensorflow as tf
from keras.layers import Input, Dense, Lambda, Flatten, Reshape, Activation, Dropout, Add, TimeDistributed, Multiply, Conv1D, Conv2D, MaxPooling1D, AveragePooling1D
from keras.models import Model, Sequential, load_model
from keras import backend as K
from keras import metrics
from keras import optimizers
from keras.callbacks import History, ModelCheckpoint
class WaveNetClassifier():
def __init__(self, input_shape, output_shape, kernel_size = 2, dilation_depth = 9, n_filters = 40, task = 'classification', regression_range = None, load=False, load_dir='./'):
"""
Parameters:
input_shape: (tuple) tuple of input shape. (e.g. If input is 6s raw waveform with sampling rate = 16kHz, (96000,) is the input_shape)
output_shape: (tuple)tuple of output shape. (e.g. If we want classify the signal into 100 classes, (100,) is the output_shape)
kernel_size: (integer) kernel size of convolution operations in residual blocks
dilation_depth: (integer) type total depth of residual blocks
n_filters: (integer) # of filters of convolution operations in residual blocks
task: (string) 'classification' or 'regression'
regression_range: (list or tuple) target range of regression task
load: (bool) load previous WaveNetClassifier or not
load_dir: (string) the directory where the previous model exists
"""
# save task info
if task == 'regression':
if regression_range[0] == 0:
self.activation = 'sigmoid'
self.scale_ratio = regression_range[1]
elif regression_range[0] == - regression_range[1]:
self.activation = 'tanh'
self.scale_ratio = regression_range[1]
elif regression_range == None:
self.activation = 'linear'
self.scale_ratio = 1
else:
print('ERROR: wrong regression range')
sys.exit()
elif task == 'classification':
self.activation = 'softmax'
self.scale_ratio = 1
else:
print('ERROR: wrong task')
sys.exit()
# save input info
if len(input_shape) == 1:
self.expand_dims = True
elif len(input_shape) == 2:
self.expand_dims = False
else:
print('ERROR: wrong input shape')
sys.exit()
self.input_shape = input_shape
# save output info
if len(output_shape) == 1:
self.time_distributed = False
elif len(output_shape) == 2:
self.time_distributed = True
else:
print('ERROR: wrong output shape')
sys.exit()
self.output_shape = output_shape
# save hyperparameters of WaveNet
self.kernel_size = kernel_size
self.dilation_depth = dilation_depth
self.n_filters = n_filters
self.manual_loss = None
if load is True:
self.model = load_model(load_dir+"saved_wavenet_clasifier.h5", custom_objects={'tf':tf})
self.prev_history = pd.read_csv(load_dir+'wavenet_classifier_training_history.csv')
self.start_idx = len(self.prev_history)
self.history = None
else:
self.model = self.construct_model()
self.start_idx = 0
self.history = None
self.prev_history = None
def residual_block(self, x, i):
tanh_out = Conv1D(self.n_filters,
self.kernel_size,
dilation_rate = self.kernel_size**i,
padding='causal',
name='dilated_conv_%d_tanh' % (self.kernel_size ** i),
activation='tanh'
)(x)
sigm_out = Conv1D(self.n_filters,
self.kernel_size,
dilation_rate = self.kernel_size**i,
padding='causal',
name='dilated_conv_%d_sigm' % (self.kernel_size ** i),
activation='sigmoid'
)(x)
z = Multiply(name='gated_activation_%d' % (i))([tanh_out, sigm_out])
skip = Conv1D(self.n_filters, 1, name='skip_%d'%(i))(z)
res = Add(name='residual_block_%d' % (i))([skip, x])
return res, skip
def construct_model(self):
x = Input(shape=self.input_shape, name='original_input')
if self.expand_dims == True:
x_reshaped = Reshape(self.input_shape + (1,), name='reshaped_input')(x)
else:
x_reshaped = x
skip_connections = []
out = Conv1D(self.n_filters, 2, dilation_rate=1, padding='causal', name='dilated_conv_1')(x_reshaped)
for i in range(1, self.dilation_depth + 1):
out, skip = self.residual_block(out,i)
skip_connections.append(skip)
out = Add(name='skip_connections')(skip_connections)
out = Activation('relu')(out)
out = Conv1D(self.n_filters, 80, strides = 1, padding='same', name='conv_5ms', activation = 'relu')(out)
out = AveragePooling1D(80, padding='same', name='downsample_to_200Hz')(out)
if self.time_distributed:
target_kernel_size = (int) (self.input_shape[0] / 80 / self.output_shape[0]) # prev_len / x = target_len => x = prev_len / target_len
out = Conv1D(self.n_filters, target_kernel_size, padding='same', name = 'conv_fit_to_target', activation='relu')(out)
out = Conv1D(self.output_shape[1], target_kernel_size, padding='same', name='conv_final')(out)
out = AveragePooling1D(target_kernel_size, padding='same')(out)
out = TimeDistributed(Activation(self.activation))(out)
else:
out = Conv1D(self.n_filters, 100, padding='same', activation='relu', name='conv_500ms')(out)
out = Conv1D(self.output_shape[0], 100, padding='same', activation='relu', name='conv_500ms_target_shape')(out)
out = AveragePooling1D(100, padding='same',name = 'downsample_to_2Hz')(out)
out = Conv1D(self.output_shape[0], (int) (self.input_shape[0] / 8000), padding='same', name='final_conv')(out)
out = AveragePooling1D((int) (self.input_shape[0] / 8000), name='final_pooling')(out)
out = Reshape(self.output_shape)(out)
out = Activation(self.activation)(out)
if self.scale_ratio != 1:
out = Lambda(lambda x: x * self.scale_ratio, name='output_reshaped')(out)
model = Model(x, out)
model.summary()
return model
def get_model(self):
return self.model
def add_loss(self, loss):
self.manual_loss = loss
def fit(self, X, Y, validation_data = None, epochs = 100, batch_size = 32, optimizer='adam', save=False, save_dir='./'):
# set default losses if not defined
if self.manual_loss is not None:
loss = self.manual_loss
metrics = None
else:
if self.task == 'classification':
loss = 'categorical_crossentropy'
metrics = ['accuracy']
else:
loss = 'mean_squared_error'
metrics = None
# set callback functions
if save:
saved = save_dir + "saved_wavenet_clasifier.h5"
hist = save_dir + 'wavenet_classifier_training_history.csv'
if validation_data is None:
checkpointer = ModelCheckpoint(filepath=saved, monitor='loss', verbose=1, save_best_only=True)
else:
checkpointer = ModelCheckpoint(filepath=saved, monitor='val_loss', verbose=1, save_best_only=True)
history = History()
callbacks = [history, checkpointer]
else:
callbacks = None
# compile the model
self.model.compile(optimizer, loss, metrics)
try:
self.history = self.model.fit(X, Y, shuffle = True, batch_size=batch_size, epochs = epochs, validation_data = validation_data, callbacks=callbacks, initial_epoch=self.start_idx)
except:
if save:
df = pd.DataFrame.from_dict(history.history)
df.to_csv(hist, encoding='utf-8', index=False)
raise
sys.exit()
return self.history
def predict(self, x):
return self.model.predict(x)