-
Notifications
You must be signed in to change notification settings - Fork 10
/
preprocessing.py
220 lines (177 loc) · 8.65 KB
/
preprocessing.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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
import numpy as np
import pandas as pd
from sklearn.base import TransformerMixin
from sklearn.impute import SimpleImputer
from sklearn.preprocessing import MinMaxScaler, RobustScaler, StandardScaler
def split_sequence(sequence, look_back_window: int, forecast_horizon: int, stride: int = 1):
X, y = [], []
for i in range(0, len(sequence), stride):
# find the end x and y
end_ix = i + look_back_window
end_iy = end_ix + forecast_horizon
# check if there is enough elements to fill this x, y pair
if end_iy > len(sequence):
break
X.append(sequence[i:end_ix])
y.append(sequence[end_iy - 1 if forecast_horizon == 1 else end_ix:end_iy])
return np.asarray(X), np.asarray(y)
class TimeSeriesImputer(TransformerMixin):
def __init__(self, method: str = 'linear', fail_save: TransformerMixin = SimpleImputer()):
self.method = method
self.fail_save = fail_save
def fit(self, data):
if self.fail_save:
self.fail_save.fit(data)
return self
def transform(self, data):
# Interpolate missing values in columns
if not isinstance(data, pd.DataFrame):
data = pd.DataFrame(data)
data = data.interpolate(method=self.method, limit_direction='both')
# spline or time may be better?
if self.fail_save:
data = self.fail_save.transform(data)
return data
def difference(dataset, interval=1, relative=False, min_price=1e-04):
delta = []
for i in range(interval, len(dataset)):
value = dataset[i] - dataset[i - interval]
if relative:
prev_price = dataset[i - interval]
prev_price[prev_price == 0] = min_price
value /= prev_price
delta.append(value)
return np.asarray(delta)
class TimeSeriesPreprocessor(TransformerMixin):
def __init__(self, look_back_window: int, forecast_horizon: int, stride: int, diff_order: int,
relative_diff: bool = True, splitXy: bool = True, scaling: str = 'minmax'):
if not (look_back_window > 0 and forecast_horizon > 0 and stride > 0):
raise ValueError('look_back_window, forecast_horizon and stride must be positive')
# if stride < look_back_window + forecast_horizon:
# raise PendingDeprecationWarning('Setting stride to less than look_back_window + forecast_horizon may not'
# 'be supported in the future due to potential data leak.')
super().__init__()
self.look_back_window = look_back_window
self.forecast_horizon = forecast_horizon
self.stride = stride
self.diff_order = diff_order
self.relative_diff = relative_diff
self.splitXy = splitXy
self.interpolation_imputer = TimeSeriesImputer(method='linear')
if scaling == 'minmax':
self.scaler = MinMaxScaler()
elif scaling == 'standard':
self.scaler = StandardScaler()
elif scaling == 'robust':
self.scaler = RobustScaler()
else:
raise ValueError('Invalid value for scaling')
def fit_transform(self, data, **fit_params):
# Fill missing values via interpolation
data = self.interpolation_imputer.fit_transform(data)
# Differencing
diff = np.array(data)
for d in range(1, self.diff_order + 1):
diff = difference(diff, relative=self.relative_diff)
data = np.append(data, np.pad(diff, pad_width=((d, 0), (0, 0))), axis=1)
if self.diff_order > 0:
data = data[:, diff.shape[1]:]
# Scale
# if self.diff_order < 1:
data = self.scaler.fit_transform(data)
if not self.splitXy:
return data
# Extract X, y from time series
X, y = split_sequence(data, self.look_back_window, self.forecast_horizon, self.stride)
return X, y
def transform(self, data):
# Fill missing values via interpolation
data = self.interpolation_imputer.transform(data)
# Differencing
diff = np.array(data)
for d in range(1, self.diff_order + 1):
diff = difference(diff, relative=self.relative_diff)
data = np.append(data, np.pad(diff, pad_width=((d, 0), (0, 0))), axis=1)
if self.diff_order > 0:
data = data[:, diff.shape[1]:]
# Scale
# if self.diff_order < 1:
data = self.scaler.transform(data)
if not self.splitXy:
return data
# Extract X, y
X, y = split_sequence(data, self.look_back_window, self.forecast_horizon, self.stride)
return X, y
class CryptoPreprocessor(TransformerMixin):
def __init__(self, look_back_window: int = 168, forecast_horizon: int = 24, stride: int = 1, diff_order: int = 0,
relative_diff: bool = True, splitXy: bool = True, target_coin: str = None):
if not (look_back_window > 0 and forecast_horizon > 0 and stride > 0):
raise ValueError('look_back_window, forecast_horizon and stride must be positive')
# if stride < look_back_window + forecast_horizon:
# raise PendingDeprecationWarning('Setting stride to less than look_back_window + forecast_horizon may not'
# 'be supported in the future due to potential data leak.')
super().__init__()
self.look_back_window = look_back_window
self.forecast_horizon = forecast_horizon
self.stride = stride
self.diff_order = diff_order
self.relative_diff = relative_diff
self.splitXy = splitXy
self.target_coin = target_coin
self.interpolation_imputer = TimeSeriesImputer(method='linear')
self.scaler = {}
def fit_transform(self, df, **fit_params):
# Fill missing values via interpolation
df = pd.DataFrame(self.interpolation_imputer.fit_transform(df), index=df.index, columns=df.columns)
# # Differencing
# diff = np.array(data)
# for d in range(1, self.diff_order + 1):
# diff = difference(diff, relative=self.relative_diff)
# data = np.append(data, np.pad(diff, pad_width=((d, 0), (0, 0))), axis=1)
# if self.diff_order > 0:
# data = data[:, diff.shape[1]:]
# Scale
# if self.diff_order < 1:
# todo: for each coin scale all categories by fit on price
# self.scaler = {coin: StandardScaler() for coin in df.columns.get_level_values('coin')}
# for coin in self.scaler.keys():
# df.loc[:, ('price', coin)] = self.scaler[coin].fit_transform(df)
# for
# df = pd.DataFrame(self.scaler.fit_transform(df), index=df.index, columns=df.columns)
# Remove market trend from
df = df.sort_values(by=['category'], axis=1)
other_coins = [col for col in df.columns.levels[1].unique() if col != 'ETH']
df_mean = pd.DataFrame.copy(df.loc[:, (df.columns.levels[0], self.target_coin)])
for cat in df.columns.levels[0].unique():
df_mean.loc[:, (cat, self.target_coin)] = df.loc[:, (cat, other_coins)].mean(axis=1)
df = df.loc[:, (df.columns.levels[0], self.target_coin)] - df_mean
if not self.splitXy:
return df
# Extract X, y from time series
X, y = split_sequence(df, self.look_back_window, self.forecast_horizon, self.stride)
return X, y
def transform(self, df):
# Fill missing values via interpolation
df = pd.DataFrame(self.interpolation_imputer.transform(df), index=df.index, columns=df.columns)
# # Differencing
# diff = np.array(data)
# for d in range(1, self.diff_order + 1):
# diff = difference(diff, relative=self.relative_diff)
# data = np.append(data, np.pad(diff, pad_width=((d, 0), (0, 0))), axis=1)
# if self.diff_order > 0:
# data = data[:, diff.shape[1]:]
# Scale
# if self.diff_order < 1:
df = pd.DataFrame(self.scaler.transform(df), index=df.index, columns=df.columns)
# Remove market trend from
df = df.sort_values(by=['category'], axis=1)
other_coins = [col for col in df.columns.levels[1].unique() if col != 'ETH']
df_mean = pd.DataFrame.copy(df.loc[:, (df.columns.levels[0], self.target_coin)])
for cat in df.columns.levels[0].unique():
df_mean.loc[:, (cat, self.target_coin)] = df.loc[:, (cat, other_coins)].mean(axis=1)
df = df.loc[:, (df.columns.levels[0], self.target_coin)] - df_mean
if not self.splitXy:
return df
# Extract X, y
X, y = split_sequence(df, self.look_back_window, self.forecast_horizon, self.stride)
return X, y