-
Notifications
You must be signed in to change notification settings - Fork 0
/
data_augmentation.py
143 lines (111 loc) · 5.94 KB
/
data_augmentation.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
# todo: add documentation
import json
from copy import deepcopy
import numpy as np
import random
from tqdm import tqdm
from numpy.typing import ArrayLike
from sklearn.preprocessing import MinMaxScaler
# Adapted https://maddevs.io/writeups/basic-data-augmentation-method-applied-to-time-series/
def add_gaussian_noise(time_series: list[np.ndarray], mean: float = 0.0, std_factor: float = 0.1) -> list:
"""
Adds Gaussian noise to a time series.
:param time_series: A time series to which noise is added, array-like.
:param mean: The average value of the noise, float. Default is 0.0.
:param std_factor: Standard deviation of noise, float.
The std of the signal is multiplied with it to get the std of the noise. Default is 0.1.
:return noisy_series: List of time series with added noise.
"""
noisy_series = []
for i in range(len(time_series)):
std = np.std(time_series[i])
noise = np.random.normal(mean, std * std_factor, len(time_series[i]))
noisy_series.append((time_series[i] + noise).tolist())
return noisy_series
# Adapted from https://maddevs.io/writeups/basic-data-augmentation-method-applied-to-time-series/
def time_warping(time_series: list[np.ndarray], labels: np.ndarray, num_operations: int = 50)\
-> tuple[list, list]:
"""
Applying time warping to a time series.
:param time_series: Time series, numpy array.
:param labels: Corresponding labels, numpy array.
:param num_operations: Number of insert/delete operations, int.
:return: Distorted time series.
"""
warped_series = time_series.copy()
warped_labels = labels.copy()
for _ in range(num_operations):
operation_type = random.choice(["insert", "delete"])
index = random.randint(1, len(warped_series) - 2)
if operation_type == "insert":
insert_points_num = random.randint(1, 20)
for feature_i in range(len(warped_series)):
for j in range(insert_points_num):
# Insert a value by interpolating between two adjacent points
insertion_value = (warped_series[feature_i][index + j - 1] + warped_series[feature_i][index + j]) * 0.5
warped_series[feature_i] = np.insert(warped_series[feature_i], index, insertion_value)
for j in range(insert_points_num):
# Insert a previous label
insertion_value = warped_labels[index + j - 1]
warped_labels = np.insert(warped_labels, index, insertion_value)
elif operation_type == "delete":
delete_points_num = random.randint(1, 20)
# Remove random points
for feature_i in range(len(warped_series)):
for j in range(delete_points_num):
warped_series[feature_i] = np.delete(warped_series[feature_i], index + j)
for j in range(delete_points_num):
# Delete label
warped_labels = np.delete(warped_labels, index + j)
else:
raise ValueError("Invalid operation type")
return np.array(warped_series).tolist(), warped_labels.tolist()
def apply_random_augmentation(time_series: list[np.ndarray], labels: np.ndarray) -> tuple[list, list]:
transform_type = random.choice(["noise", "time-warping", "both"])
if transform_type == "noise":
transformed_signals = add_gaussian_noise(time_series)
return transformed_signals, labels.tolist()
if transform_type == "time-warping":
# for time-warping the labels should change too
transformed_signals, transformed_labels = time_warping(time_series, labels)
return transformed_signals, transformed_labels
if transform_type == "both":
transformed_signals, transformed_labels = time_warping(time_series, labels)
transformed_signals = add_gaussian_noise(transformed_signals)
return transformed_signals, transformed_labels
def normalize_signal(sig: ArrayLike) -> list:
scaler = MinMaxScaler(feature_range=(0, 1))
sig_norm = scaler.fit_transform([[x] for x in sig]).reshape(len(sig))
return sig_norm.tolist()
def augment_data(data: dict[str, list], augment_percentage: float = 0.3, save_to_file: str | None = None) -> dict[str, list]:
augmented_data = deepcopy(data)
dataset_size = len(data['ppg_data'])
signals_to_augment = random.sample(list(np.arange(dataset_size)), int(dataset_size * augment_percentage))
for i in tqdm(signals_to_augment):
signals = [np.array(data['ppg_not_normalized'][i]),
np.array(data['hr_not_normalized'][i]),
np.array(data['accelerometer_data'][i])]
augmented_signals, augmented_label = apply_random_augmentation(signals, np.array(data['targets'][i]))
augmented_data['ppg_not_normalized'].append(augmented_signals[0])
augmented_data['hr_not_normalized'].append(augmented_signals[1])
normalized_ppg = normalize_signal(augmented_signals[0])
normalized_hr = normalize_signal(augmented_signals[1])
normalized_acc = normalize_signal(augmented_signals[2])
augmented_data['ppg_data'].append(normalized_ppg)
augmented_data['hr_data'].append(normalized_hr)
augmented_data['accelerometer_data'].append(normalized_acc)
augmented_data['targets'].append(augmented_label)
if save_to_file is not None:
with open(save_to_file, 'w') as f:
json.dump(augmented_data, f, indent=4)
return augmented_data
if __name__ == '__main__':
from utils import load_data
from os.path import join
data = load_data()
path_to_save = join('data', 'W_AUGMENTED_DATA.json')
augmented_data = augment_data(data, save_to_file=path_to_save)
print(f"Successfully created dataset with augmentation! "
f"Initial dataset size: {len(data['ppg_data'])}, augmented dataset size: {len(augmented_data['ppg_data'])}")
if path_to_save:
print(f'Augmented dataset saved to {path_to_save}.')