-
Notifications
You must be signed in to change notification settings - Fork 1
/
gcloud_training_test.py
152 lines (128 loc) · 5.67 KB
/
gcloud_training_test.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
import numpy as np
from scipy.io import wavfile
import pandas as pd
from scipy.signal import decimate
from sklearn.model_selection import train_test_split
from keras.models import Sequential
from keras.layers import Conv1D, MaxPool1D, GlobalAvgPool1D, Dropout, BatchNormalization, Dense
from keras.optimizers import Adam
from keras.callbacks import ModelCheckpoint, LearningRateScheduler, EarlyStopping
from keras.utils import np_utils
from keras.regularizers import l2
INPUT_LIB = 'input/'
SAMPLE_RATE = 44100
CLASSES = ['artifact', 'normal', 'murmur']
CODE_BOOK = {x:i for i,x in enumerate(CLASSES)}
NB_CLASSES = len(CLASSES)
def clean_filename(fname, string):
file_name = fname.split('/')[1]
if file_name[:2] == '__':
file_name = string + file_name
return file_name
def load_wav_file(name, path):
_, b = wavfile.read(path + name)
assert _ == SAMPLE_RATE
return b
def repeat_to_length(arr, length):
"""Repeats the numpy 1D array to given length, and makes datatype float"""
result = np.empty((length, ), dtype = 'float32')
l = len(arr)
pos = 0
while pos + l <= length:
result[pos:pos+l] = arr
pos += l
if pos < length:
result[pos:length] = arr[:length-pos]
return result
df = pd.read_csv(INPUT_LIB + 'set_a.csv')
df['fname'] = df['fname'].apply(clean_filename, string='Aunlabelledtest')
df['label'].fillna('unclassified')
df['time_series'] = df['fname'].apply(load_wav_file, path=INPUT_LIB + 'set_a/')
df['len_series'] = df['time_series'].apply(len)
MAX_LEN = max(df['len_series'])
df['time_series'] = df['time_series'].apply(repeat_to_length, length=MAX_LEN)
x_data = np.stack(df['time_series'].values, axis=0)
new_labels =[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 1, 1, 2, 1,
2, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2, 2, 2, 2, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 2, 1, 1,
1, 1, 1, 1, 1, 2, 2, 1, 1, 1, 1, 1, 2, 1, 0, 2, 2, 1, 1, 1, 1, 1,
0, 1, 0, 1, 1, 1, 2, 1, 0, 1, 1, 1, 1, 1, 2, 0, 0, 0, 0, 0, 0, 0,
1, 0, 0, 0, 0, 0, 0, 1, 0, 2, 1, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2]
new_labels = np.array(new_labels, dtype='int')
y_data = np_utils.to_categorical(new_labels)
x_train, x_test, y_train, y_test, train_filenames, test_filenames = \
train_test_split(x_data, y_data, df['fname'].values, test_size=0.25)
x_train = decimate(x_train, 8, axis=1, zero_phase=True)
x_train = decimate(x_train, 8, axis=1, zero_phase=True)
x_train = decimate(x_train, 4, axis=1, zero_phase=True)
x_test = decimate(x_test, 8, axis=1, zero_phase=True)
x_test = decimate(x_test, 8, axis=1, zero_phase=True)
x_test = decimate(x_test, 4, axis=1, zero_phase=True)
x_train = x_train / np.std(x_train, axis=1).reshape(-1,1)
x_test = x_test / np.std(x_test, axis=1).reshape(-1,1)
x_train = x_train[:,:,np.newaxis]
x_test = x_test[:,:,np.newaxis]
model = Sequential()
model.add(Conv1D(filters=4, kernel_size=9, activation='relu',
input_shape = x_train.shape[1:],
kernel_regularizer = l2(0.025)))
model.add(MaxPool1D(strides=4))
model.add(BatchNormalization())
model.add(Conv1D(filters=4, kernel_size=9, activation='relu',
kernel_regularizer = l2(0.05)))
model.add(MaxPool1D(strides=4))
model.add(BatchNormalization())
model.add(Conv1D(filters=8, kernel_size=9, activation='relu',
kernel_regularizer = l2(0.1)))
model.add(MaxPool1D(strides=4))
model.add(BatchNormalization())
model.add(Conv1D(filters=16, kernel_size=9, activation='relu'))
model.add(MaxPool1D(strides=4))
model.add(BatchNormalization())
model.add(Dropout(0.25))
model.add(Conv1D(filters=64, kernel_size=4, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.5))
model.add(Conv1D(filters=32, kernel_size=1, activation='relu'))
model.add(BatchNormalization())
model.add(Dropout(0.75))
model.add(GlobalAvgPool1D())
model.add(Dense(3, activation='softmax'))
def batch_generator(x_train, y_train, batch_size):
"""
Rotates the time series randomly in time
"""
x_batch = np.empty((batch_size, x_train.shape[1], x_train.shape[2]), dtype='float32')
y_batch = np.empty((batch_size, y_train.shape[1]), dtype='float32')
full_idx = range(x_train.shape[0])
while True:
batch_idx = np.random.choice(full_idx, batch_size)
x_batch = x_train[batch_idx]
y_batch = y_train[batch_idx]
for i in range(batch_size):
sz = np.random.randint(x_batch.shape[1])
x_batch[i] = np.roll(x_batch[i], sz, axis = 0)
yield x_batch, y_batch
weight_saver = ModelCheckpoint('set_a_weights.h5', monitor='val_loss',
save_best_only=True, save_weights_only=True)
model.compile(optimizer=Adam(1e-4), loss='categorical_crossentropy', metrics=['accuracy'])
annealer = LearningRateScheduler(lambda x: 1e-3 * 0.8**x)
hist = model.fit_generator(batch_generator(x_train, y_train, 8),
epochs=1000, steps_per_epoch=1000,
validation_data=(x_test, y_test),
callbacks=[weight_saver, annealer],
verbose=2)
model.load_weights('set_a_weights.h5')
y_hat = model.predict(x_test)
np.set_printoptions(precision=2, suppress=True)
for i in range(3):
print(CLASSES[i])
y_pred = np.argmax(y_hat, axis=1)
y_true = np.argmax(y_test, axis=1)
for i in range(len(y_true)):
if y_pred[i] != y_true[i]:
print("File: {}, Pred: {}, True: {}".format(
test_filenames[i],
CLASSES[y_pred[i]], CLASSES[y_true[i]]))