-
Notifications
You must be signed in to change notification settings - Fork 89
/
augmenter.py
199 lines (166 loc) · 6.48 KB
/
augmenter.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
# -*- coding: utf-8 -*-
# copyright: aeon developers, BSD-3-Clause License (see LICENSE file)
"""Series transformers for time series augmentation."""
__author__ = ["MrPr3ntice", "MFehsenfeld", "iljamaurer"]
__all__ = [
"WhiteNoiseAugmenter",
"ReverseAugmenter",
"InvertAugmenter",
"RandomSamplesAugmenter",
]
import numpy as np
import pandas as pd
from scipy.stats import norm
from sklearn.utils import check_random_state
from aeon.transformations.base import BaseTransformer
class _AugmenterTags:
_tags = {
"scitype:transform-input": "Series",
"scitype:transform-output": "Series",
"scitype:transform-labels": "None",
"scitype:instancewise": True,
"capability:missing_values": False,
"y_inner_mtype": "pd.Series",
"X_inner_mtype": "pd.DataFrame",
"X-y-must-have-same-index": False,
"fit_is_empty": True,
"transform-returns-same-time-index": False,
"capability:inverse_transform": False,
}
class WhiteNoiseAugmenter(_AugmenterTags, BaseTransformer):
r"""Augmenter adding Gaussian (i.e. white) noise to the time series.
If `transform` is given time series :math:`X={x_1, x_2, ... , x_n}`, then
returns :math:`X_t={x_1+e_1, x_2+e_2, ..., x_n+e_n}` where :math:`e_i` are
i.i.d. random draws from a normal distribution with mean :math:`\mu` = 0
and standard deviation :math:`\sigma` = ``scale``.
Time series augmentation by adding Gaussian Noise has been discussed among
others in [1] and [2].
Parameters
----------
scale: float, default=1.0
Scale parameter specifies the standard deviation.
random_state: None or int or ``np.random.RandomState`` instance, default = None
If int or RandomState, use it for drawing the random variates.
If None, rely on ``self.random_state``.
References
----------
..[1]: WEN, Qingsong, et al. Time series data augmentation for deep
learning: A survey. arXiv preprint arXiv:2002.12478, 2020.
..[2]: IWANA, Brian Kenji; UCHIDA, Seiichi. An empirical survey of data
augmentation for time series classification with neural networks. Plos
one, 2021, 16. Jg., Nr. 7, S. e0254841.
"""
_allowed_statistics = [np.std]
def __init__(self, scale=1.0, random_state=42):
self.scale = scale
self.random_state = random_state
super().__init__()
def _transform(self, X, y=None):
if self.scale in self._allowed_statistics:
scale = self.scale(X)
elif isinstance(self.scale, (int, float)):
scale = self.scale
else:
raise TypeError(
"Type of parameter 'scale' must be a non-negative float value."
)
return X[0] + norm.rvs(0, scale, size=len(X), random_state=self.random_state)
class ReverseAugmenter(_AugmenterTags, BaseTransformer):
r"""Augmenter reversing the time series.
If `transform` is given a time series :math:`X={x_1, x_2, ... , x_n}`, then
returns :math:`X_t={x_n, x_{n-1}, ..., x_2, x_1}`.
Time series augmentation by reversing has been discussed e.g. in [1].
Examples
--------
>>> X = pd.Series([1,2,3,4,5])
>>> augmenter = ReverseAugmenter()
>>> Xt = augmenter.fit_transform(X)
>>> Xt
0 5
1 4
2 3
3 2
4 1
dtype: int64
References
----------
..[1]: IWANA, Brian Kenji; UCHIDA, Seiichi. An empirical survey of data
augmentation for time series classification with neural networks. Plos
one, 2021, 16. Jg., Nr. 7, S. e0254841.
"""
def __init__(self):
super().__init__()
def _transform(self, X, y=None):
return X.loc[::-1].reset_index(drop=True, inplace=False)
class InvertAugmenter(_AugmenterTags, BaseTransformer):
r"""Augmenter inverting the time series by multiplying it by -1.
If `transform` is given a time series :math:`X={x_1, x_2, ... , x_n}`, then
returns :math:`X_t={-x_1, -x_2, ... , -x_n}`.
Examples
--------
>>> X = pd.Series([1,2,3,4,5])
>>> augmenter = InvertAugmenter()
>>> Xt = augmenter.fit_transform(X)
>>> Xt
0 -1
1 -2
2 -3
3 -4
4 -5
dtype: int64
"""
def __init__(self):
super().__init__()
def _transform(self, X, y=None):
return X.mul(-1)
class RandomSamplesAugmenter(_AugmenterTags, BaseTransformer):
r"""Draw random samples from time series.
`transform` takes a time series :math:`X={x_1, x_2, ... , x_m}` with :math:`m`
elements and returns :math:`X_t={x_i, x_{i+1}, ... , x_n}`, where
:math:`{x_i, x_{i+1}, ... , x_n}` are :math:`n`=``n`` random samples drawn
from :math:`X` (with or `without_replacement`).
Parameters
----------
n : int or float, default = 1.0
Specify an exact number of samples to draw, set `n` to an int value.
To specify the returned samples as a proportion of the given times series
set `n` to a float value :math:`n \in [0, 1]`. By default, the same number of
samples is returned as given by the input time series.
without_replacement : bool, default = True
Whether to draw without replacement. If True, every sample of the input times
series `X` will appear at most once in `Xt`.
random_state: None or int or ``np.random.RandomState`` instance, default = None
If int or RandomState, use it for drawing the random variates.
If None, rely on ``self.random_state``.
"""
def __init__(
self,
n=1.0,
without_replacement=True,
random_state=42,
):
if isinstance(n, float):
if n <= 0.0 or not np.isfinite(n):
raise ValueError("n must be a positive, finite number.")
elif isinstance(n, int):
if n < 1 or not np.isfinite(n):
raise ValueError("n must be a finite number >= 1.")
else:
raise ValueError("n must be int or float, not " + str(type(n))) + "."
self.n = n
self.without_replacement = without_replacement
self.random_state = random_state
super().__init__()
def _transform(self, X, y=None):
if isinstance(self.n, float):
n = int(np.ceil(self.n * len(X)))
else:
n = self.n
rng = check_random_state(self.random_state)
values = np.concatenate(X.values)
if self.without_replacement:
replace = False
else:
replace = True
Xt = rng.choice(values, n, replace)
return pd.DataFrame(Xt)