Newer
Older
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
def masktraining_skip_2chan(path_train, path_valid, batch_size, epochs, path_save_model, path_weights_model, plot_name, option, reduction_divisor):
# initializing
#!/usr/bin/env python3
# -*- coding: utf8 -*-
import ctypes
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '1'
os.environ['TF_ENABLE_XLA'] = '1'
os.environ['TF_ENABLE_AUTO_MIXED_PRECISION'] = '1'
os.environ['TF_ENABLE_CUDNN_RNN_TENSOR_OP_MATH_FP32'] = '1'
os.environ['TF_DISABLE_CUDNN_TENSOR_OP_MATH'] = '1'
os.environ['TF_ENABLE_CUBLAS_TENSOR_OP_MATH_FP32'] = '1'
import numpy as np
import tensorflow as tf
from tensorflow import keras
from keras.layers import Add, Multiply, Input, Dense, Flatten, Dropout, Conv2D, MaxPooling2D, Conv2DTranspose, LeakyReLU, Reshape, Activation, BatchNormalization, UpSampling2D
from keras.models import Model
from keras.constraints import max_norm
from keras.optimizers import Adam, SGD
from keras.utils import to_categorical, normalize
from keras.callbacks import ModelCheckpoint
from keras import losses
from keras import backend as K
import matplotlib.pyplot as plt
import random
# checking for gpus and using it/them
gpus = tf.config.experimental.list_physical_devices('GPU')
if gpus:
try:
tf.config.experimental.set_virtual_device_configuration(gpus[0], [tf.config.experimental.VirtualDeviceConfiguration(memory_limit=4096)])
except RuntimeError as e:
print(e)
# custom generator import
from Generators.MaskGenerator3_realimag import DataGenerator
# ERSETZEN!
def count_files(dir):
return len([1 for x in list(os.scandir(dir)) if x.is_file()])
len_train = count_files(path_train)
len_valid = count_files(path_valid)
# generators for train and validation data
train_generator = DataGenerator(path_train, option, reduction_divisor, len_train, batch_size, True)
valid_generator = DataGenerator(path_valid, option, reduction_divisor, len_valid, batch_size, True)
# building model
def build_model():
max_norm_value = 100.0
# for first channel
# defining inputs and normalizing just noisy input
input_noisy_1 = Input(shape=(260,5,1))
input_noise_1 = Input(shape=(260,1,1))
input_speech_1 = Input(shape=(260,1,1))
normalized_noisy_1 = BatchNormalization()(input_noisy_1)
# normalized_noise = BatchNormalization()(input_noise)
# normalized_speech = BatchNormalization()(input_speech)
# encoder of net
conv_1_1 = Conv2D(filters=32, kernel_size=(1,3),strides=1,padding="valid",kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(normalized_noisy_1)
leakyrelu_1_1 = LeakyReLU()(conv_1_1)
conv_1_2 = Conv2D(filters=32, kernel_size=(1,3),strides=1,padding="valid",kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_1_1)
leakyrelu_1_2 = LeakyReLU()(conv_1_2)
conv_1_3 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_1_2)
leakyrelu_1_3 = LeakyReLU()(conv_1_3)
maxpool_1_1 = MaxPooling2D(pool_size=(2,1))(leakyrelu_1_3)
conv_1_4 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(maxpool_1_1)
leakyrelu_1_4 = LeakyReLU()(conv_1_4)
maxpool_1_2 = MaxPooling2D(pool_size=(2,1))(leakyrelu_1_4)
conv_1_5 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(maxpool_1_2)
leakyrelu_1_5 = LeakyReLU()(conv_1_5)
conv_1_6 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_1_5)
leakyrelu_1_6 = LeakyReLU()(conv_1_6)
# decoder of Net
convtrans_1_1 = Conv2DTranspose(filters=32, kernel_size=(16,1),strides=1,padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_1_6)
leakyrelu_1_7 = LeakyReLU()(convtrans_1_1)
convtrans_1_2 = Conv2DTranspose(filters=32, kernel_size=(16,1),strides=1, padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_1_7)
leakyrelu_1_8 = LeakyReLU()(convtrans_1_2)
skip_1_1 = Add()([maxpool_1_2,leakyrelu_1_8])
up_1_1 = UpSampling2D(size=(2,1))(skip_1_1)
conv_1_7 = Conv2D(filters=32, kernel_size=(16,1), strides=1, padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(up_1_1)
leakyrelu_1_9 = LeakyReLU()(conv_1_7)
skip_1_2 = Add()([leakyrelu_1_4,leakyrelu_1_9])
up_1_2 = UpSampling2D(size=(2,1))(skip_1_2)
conv_1_8 = Conv2D(filters=32, kernel_size=(16,1), strides=1, padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(up_1_2)
leakyrelu_1_10 = LeakyReLU()(conv_1_8)
skip_1_3 = Add()([leakyrelu_1_3,leakyrelu_1_10])
# mask from noisy input
mask_1 = Conv2D(filters=1, kernel_size=(16,1),strides=1, padding="same", activation='sigmoid', kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(skip_1_3)
# filtered speech and noise component
n_tilde_1 = Multiply()([mask_1,input_noise_1])
s_tilde_1 = Multiply()([mask_1,input_speech_1])
# for secondchannel
# defining inputs and normalizing just noisy input
input_noisy_2 = Input(shape=(260,5,1))
input_noise_2 = Input(shape=(260,1,1))
input_speech_2 = Input(shape=(260,1,1))
normalized_noisy_2 = BatchNormalization()(input_noisy_2)
# normalized_noise = BatchNormalization()(input_noise)
# normalized_speech = BatchNormalization()(input_speech)
# encoder of net
conv_2_1 = Conv2D(filters=32, kernel_size=(1,3),strides=1,padding="valid",kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(normalized_noisy_2)
leakyrelu_2_1 = LeakyReLU()(conv_2_1)
conv_2_2 = Conv2D(filters=32, kernel_size=(1,3),strides=1,padding="valid",kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_2_1)
leakyrelu_2_2 = LeakyReLU()(conv_2_2)
conv_2_3 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_2_2)
leakyrelu_2_3 = LeakyReLU()(conv_2_3)
maxpool_2_1 = MaxPooling2D(pool_size=(2,1))(leakyrelu_2_3)
conv_2_4 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(maxpool_2_1)
leakyrelu_2_4 = LeakyReLU()(conv_2_4)
maxpool_2_2 = MaxPooling2D(pool_size=(2,1))(leakyrelu_2_4)
conv_2_5 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(maxpool_2_2)
leakyrelu_2_5 = LeakyReLU()(conv_2_5)
conv_2_6 = Conv2D(filters=32, kernel_size=(16,1),strides=1,padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_2_5)
leakyrelu_2_6 = LeakyReLU()(conv_2_6)
# decoder of Net
convtrans_2_1 = Conv2DTranspose(filters=32, kernel_size=(16,1),strides=1,padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_2_6)
leakyrelu_2_7 = LeakyReLU()(convtrans_2_1)
convtrans_2_2 = Conv2DTranspose(filters=32, kernel_size=(16,1),strides=1, padding="valid", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(leakyrelu_2_7)
leakyrelu_2_8 = LeakyReLU()(convtrans_2_2)
skip_2_1 = Add()([maxpool_2_2,leakyrelu_2_8])
up_2_1 = UpSampling2D(size=(2,1))(skip_2_1)
conv_2_7 = Conv2D(filters=32, kernel_size=(16,1), strides=1, padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(up_2_1)
leakyrelu_2_9 = LeakyReLU()(conv_2_7)
skip_2_2 = Add()([leakyrelu_2_4,leakyrelu_2_9])
up_2_2 = UpSampling2D(size=(2,1))(skip_2_2)
conv_2_8 = Conv2D(filters=32, kernel_size=(16,1), strides=1, padding="same", kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(up_2_2)
leakyrelu_2_10 = LeakyReLU()(conv_2_8)
skip_2_3 = Add()([leakyrelu_2_3,leakyrelu_2_10])
# mask from noisy input
mask_2 = Conv2D(filters=1, kernel_size=(16,1),strides=1, padding="same", activation='sigmoid', kernel_constraint=max_norm(max_norm_value), kernel_initializer="he_normal")(skip_2_3)
# filtered speech and noise component
n_tilde_2 = Multiply()([mask_2,input_noise_2])
s_tilde_2 = Multiply()([mask_2,input_speech_2])
# defining model
model = Model(inputs=[input_noisy_1,input_noise_1,input_speech_1,input_noisy_2,input_noise_2,input_speech_2], outputs=[n_tilde_1,s_tilde_1,n_tilde_2,s_tilde_2])
return model
# if __name__ == "__main__":
# build model and compile it
model = build_model()
model.compile(loss='mse', loss_weights=[0.25, 0.25, 0.25, 0.25], optimizer='adam', metrics=['mape','accuracy'])
# making directory for weights
if not os.path.exists(path_weights_model):
try:
os.makedirs(path_weights_model)
except Exception as e:
print("Konnte Ordner fuer Gewichte nicht erstellen" + str(e))
# defining how weights are saved
filepath = path_weights_model + "/weights-{epoch:04d}.hdf5"
checkpoint = ModelCheckpoint(
filepath,
monitor='loss',
verbose=0,
save_best_only=True,
mode='min'
)
model.summary()
# train model
history = model.fit(train_generator, steps_per_epoch = int(np.floor(len_train // reduction_divisor / batch_size)), epochs = epochs, validation_data = valid_generator, validation_steps = int(np.floor(len_valid // reduction_divisor / batch_size)), callbacks=[checkpoint], use_multiprocessing=True)
model.summary()
# save model after training
model.save(path_save_model)
index = path_save_model.split("/")
path_save_nomodel = '/'.join(index[0:-1])
loss = history.history["loss"]
multiply_1_loss = history.history["multiply_1_loss"]
multiply_2_loss = history.history["multiply_2_loss"]
multiply_3_loss = history.history["multiply_3_loss"]
multiply_4_loss = history.history["multiply_4_loss"]
multiply_1_mape = history.history["multiply_1_mape"]
multiply_2_mape = history.history["multiply_2_mape"]
multiply_3_mape = history.history["multiply_3_mape"]
multiply_4_mape = history.history["multiply_4_mape"]
multiply_1_accuracy = history.history["multiply_1_accuracy"]
multiply_2_accuracy = history.history["multiply_2_accuracy"]
multiply_3_accuracy = history.history["multiply_3_accuracy"]
multiply_4_accuracy = history.history["multiply_4_accuracy"]
val_loss = history.history["val_loss"]
val_multiply_1_loss = history.history["val_multiply_1_loss"]
val_multiply_2_loss = history.history["val_multiply_2_loss"]
val_multiply_3_loss = history.history["val_multiply_3_loss"]
val_multiply_4_loss = history.history["val_multiply_4_loss"]
val_multiply_1_mape = history.history["val_multiply_1_mape"]
val_multiply_2_mape = history.history["val_multiply_2_mape"]
val_multiply_3_mape = history.history["val_multiply_3_mape"]
val_multiply_4_mape = history.history["val_multiply_4_mape"]
val_multiply_1_accuracy = history.history["val_multiply_1_accuracy"]
val_multiply_2_accuracy = history.history["val_multiply_2_accuracy"]
val_multiply_3_accuracy = history.history["val_multiply_3_accuracy"]
val_multiply_4_accuracy = history.history["val_multiply_4_accuracy"]
dict_history = {'loss':loss, 'multiply_1_loss':multiply_1_loss, 'multiply_2_loss':multiply_2_loss, 'multiply_3_loss':multiply_3_loss, 'multiply_4_loss':multiply_4_loss, 'multiply_1_mape':multiply_1_mape, 'multiply_2_mape':multiply_2_mape, 'multiply_3_mape':multiply_3_mape, 'multiply_4_mape':multiply_4_mape, 'multiply_1_accuracy':multiply_1_accuracy, 'multiply_2_accuracy':multiply_2_accuracy, 'multiply_3_accuracy':multiply_3_accuracy, 'multiply_4_accuracy':multiply_4_accuracy, 'val_loss':val_loss, 'val_multiply_1_loss':val_multiply_1_loss, 'val_multiply_2_loss':val_multiply_2_loss, 'val_multiply_3_loss':val_multiply_3_loss, 'val_multiply_4_loss':val_multiply_4_loss, 'val_multiply_1_mape':val_multiply_1_mape, 'val_multiply_2_mape':val_multiply_2_mape, 'val_multiply_3_mape':val_multiply_3_mape, 'val_multiply_4_mape':val_multiply_4_mape, 'val_multiply_1_accuracy':val_multiply_1_accuracy, 'val_multiply_2_accuracy':val_multiply_2_accuracy, 'val_multiply_3_accuracy':val_multiply_3_accuracy, 'val_multiply_4_accuracy':val_multiply_4_accuracy}
# loss_history = history.history["loss"]
# mape_history = history.history["mape"]
# accuracy_history = history.history["accuracy"]
# val_loss_history = history.history["val_loss"]
# val_mape_history = history.history["val_mape"]
# val_accuracy_history = history.history["val_accuracy"]
#
# dict_history = {'loss':loss_history, 'mape':mape_history, 'accuracy':accuracy_history, 'val_loss':val_loss_history, 'val_mape':val_mape_history, 'val_accuracy':val_accuracy_history}
np.save(path_save_nomodel + '/' + plot_name, dict_history)
# plot train loss and validation loss after training
plt.style.use("ggplot")
plt.figure()
plt.plot(history.history["loss"], label="train_loss")
plt.plot(history.history["val_loss"], label="val_loss")
plt.title("Training Loss and Accuracy")
plt.xlabel("Epochs")
plt.ylabel("Loss/Accuracy")
plt.legend(loc="lower left")
index = path_save_model.split("/")
path_save_nomodel = '/'.join(index[0:-1])
plt.savefig(path_save_nomodel + '/' + plot_name + '.png', dpi=400)
# plt.show()
return