-
Notifications
You must be signed in to change notification settings - Fork 2
/
data.py
234 lines (205 loc) · 9.89 KB
/
data.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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
import numpy as np
import random
import torch
import torch.utils.data as torchdata
import pandas as pd
class timeseries_dataset():
def __init__(self, name, dim_inputseqlen, dim_outputseqlen, seed, train_ratio=0.8):
self.name = name
self.dim_inputseqlen = dim_inputseqlen
self.dim_outputseqlen = dim_outputseqlen
self.seed = seed
self.train_ratio = train_ratio
def load(self, mode):
if self.name.startswith('same'):
output = syn_data(self.dim_inputseqlen, self.dim_outputseqlen, mode, self.name, seed=self.seed, filename=self.name, train_ratio=self.train_ratio)
else:
if self.name == 'favorita_family':
filename = 'y15_s0_family26'
folder = 'kaggle_favorita'
key = 'favorita'
elif self.name == 'favorita_family_store':
filename = 'y15_s0_store45'
folder = 'kaggle_favorita'
key = 'favorita'
elif self.name == 'traffic':
filename = 'y22_ca_s19'
folder = 'traffic'
key = 'traffic'
elif self.name == 'stock_vol':
filename = 'y2021_stock12'
folder = 'stock'
key = 'stock'
output = realworld_data(self.dim_inputseqlen, self.dim_outputseqlen, mode, self.name, seed=self.seed, filename=filename, folder=folder, key=key, train_ratio=self.train_ratio)
return output
class syn_data(torchdata.Dataset):
def __init__(self, dim_inputseqlen, dim_outputseqlen, mode, name, seed=0, filename='',train_ratio=0.8):
self.dim_inputseqlen = dim_inputseqlen
self.dim_outputseqlen = dim_outputseqlen
self.window = dim_inputseqlen + dim_outputseqlen
# self.dim_maxseqlen = dim_maxseqlen
self.mode = mode
self.name = name
self.dataset_filename = filename
# test large datasets
if filename.endswith('1m'):
self.train_maxdate = 540000
self.validate_maxdate = 720000
elif filename.endswith('100k'):
self.train_maxdate = 54000
self.validate_maxdate = 72000
elif filename.endswith('10k'):
self.train_maxdate = 5400
self.validate_maxdate = 7200
elif filename.endswith('1k'):
self.train_maxdate = 540
self.validate_maxdate = 720
else:
self.train_maxdate = 270
self.validate_maxdate = 360
self.seed = seed
self.links = None
self.train_doms = None
self.test_doms = None
self.train_ratio = train_ratio
self.X, self.Y = self.get_data()
def __len__(self):
return len(self.index)
def __getitem__(self, idx):
x = self.X[self.index[idx]:self.index[idx] + self.dim_outputseqlen + self.dim_inputseqlen]
y = self.Y[self.index[idx]:self.index[idx] + self.dim_outputseqlen + self.dim_inputseqlen]
return x, y
def get_data(self):
df = pd.read_hdf('data/syn/df_{}.h5'.format(self.dataset_filename), key=self.dataset_filename)
index = pd.read_hdf('data/syn/df_{}.h5'.format(self.dataset_filename), key='index')
df_Y = df[['x']]
df_X = df[['domain','x_lagged']]
all_doms = df_X['domain'].unique()
random.seed(self.seed)
random.shuffle(all_doms)
num_doms = len(all_doms)
num_train_doms = int(num_doms*self.train_ratio)
train_doms = all_doms[:num_train_doms]
test_doms = all_doms[num_train_doms:]
if self.mode == 'train':
print(df_X.columns)
print(f"domain: {df_X['domain'].unique().shape}")
print(f'all_doms:{all_doms} train_doms:{train_doms} {len(train_doms)} test_doms:{test_doms} {len(test_doms)}')
self.train_doms = train_doms
self.test_doms = test_doms
X = df_X.to_numpy(dtype='float32')
Y = df_Y.to_numpy(dtype='float32')
self.dim_input, self.dim_output = X.shape[-1], Y.shape[-1]
X, Y = torch.from_numpy(X), torch.from_numpy(Y)
# use training domain validation `IN SEARCH OF LOST DOMAIN GENERALIZATION`
if self.mode == 'train':
idx = index[(index['date'] <= self.train_maxdate) & (index['domain'].isin(train_doms))]['index'].to_numpy()
self.index = torch.from_numpy(idx)
elif self.mode == 'validate':
idx = index[(index['date'] <= self.validate_maxdate) & (index['date'] > self.train_maxdate) & (index['domain'].isin(train_doms))]['index'].to_numpy()
self.index = torch.from_numpy(idx)
elif self.mode == 'test':
idx = index[(index['date'] > self.validate_maxdate) & (index['domain'].isin(test_doms))]['index'].to_numpy()
self.index = torch.from_numpy(idx)
# TODO add comments
self.d_emb = 0 # number of embedding categoricals in input
self.d_cov = 0
self.d_lag = 1
self.x_dim = 1
self.y_dim = 1
self.d_dim = num_doms # number of domains
self.x_bin_dim = 0
self.x_con_dim = 1
return X, Y
class realworld_data(torchdata.Dataset):
def __init__(self, dim_inputseqlen, dim_outputseqlen, mode, name, seed=0, filename='', folder='', key='',train_ratio=0.8):
self.dim_inputseqlen = dim_inputseqlen
self.dim_outputseqlen = dim_outputseqlen
self.window = dim_inputseqlen + dim_outputseqlen
self.mode = mode
self.name = name
self.dataset_filename = filename
self.dataset_folder = folder
self.dataset_key = key
if 'family' in self.dataset_filename or 'store' in self.dataset_filename: # 2015-03-01 to 2015-12-31
self.train_maxdate = '2015-06-30' # 122/46/51
self.validate_maxdate = '2015-08-15'
elif self.dataset_filename.endswith('ca_s19'): # 2022-04-01 to 2017-11-30
self.train_maxdate = '2022-07-15'
self.validate_maxdate = '2022-08-20'
elif self.dataset_filename.endswith(('stock12')): # 2020-01-01 to 2021-5-31
self.train_maxdate = '2020-08-30'
self.validate_maxdate = '2020-11-30'
self.seed = seed
self.links = None
self.train_doms = None
self.test_doms = None
self.train_ratio = train_ratio
self.X, self.Y = self.get_data()
def __len__(self):
return len(self.index)
def __getitem__(self, idx):
x = self.X[self.index[idx]:self.index[idx] + self.dim_outputseqlen + self.dim_inputseqlen]
y = self.Y[self.index[idx]:self.index[idx] + self.dim_outputseqlen + self.dim_inputseqlen]
return x, y
def get_data(self):
df = pd.read_hdf(f'data/{self.dataset_folder}/df_{self.dataset_filename}.h5', key=self.dataset_key)
index = pd.read_hdf(f'data/{self.dataset_folder}/df_{self.dataset_filename}.h5', key='index')
if 'store' in self.dataset_filename:
df['unit_sales'] = np.log(df['unit_sales']+1)
df['unit_sales_lagged'] = np.log(df['unit_sales_lagged']+1)
df_Y = df[['unit_sales']]
df_X = df[['store_nbr','DayOfWeek_sin','DayOfWeek_cos','unit_sales_lagged']]
domain_name = 'store_nbr'
elif 'family' in self.dataset_filename:
df['unit_sales'] = np.log(df['unit_sales']+1)
df['unit_sales_lagged'] = np.log(df['unit_sales_lagged']+1)
df_Y = df[['unit_sales']]
df_X = df[['family','DayOfWeek_sin','DayOfWeek_cos','unit_sales_lagged']]
domain_name = 'family'
elif self.dataset_filename.endswith('ca_s19'):
df_Y = df[['vol']]
df_X = df[['station_id','DayOfWeek_sin','DayOfWeek_cos','vol_lagged']]
domain_name = 'station_id'
elif self.dataset_filename.endswith('stock12'):
df['vol'] = df['vol'] / 1e7
df['vol_lagged'] = df['vol_lagged'] / 1e7
df_Y = df[['vol']]
df_X = df[['Index','DayOfWeek_sin','DayOfWeek_cos','vol_lagged']]
domain_name = 'Index'
all_doms = df_X[domain_name].unique()
random.seed(self.seed)
random.shuffle(all_doms)
num_doms = len(all_doms)
num_train_doms = int(num_doms*self.train_ratio)
train_doms = all_doms[:num_train_doms]
test_doms = all_doms[num_train_doms:]
if self.mode == 'train':
print(df_X.columns)
print(f"domain: {df_X[domain_name].unique().shape}")
print(f'all_doms:{all_doms} train_doms:{train_doms} {len(train_doms)} test_doms:{test_doms} {len(test_doms)}')
self.train_doms = train_doms
self.test_doms = test_doms
X = df_X.to_numpy(dtype='float32')
Y = df_Y.to_numpy(dtype='float32')
self.dim_input, self.dim_output = X.shape[-1], Y.shape[-1]
X, Y = torch.from_numpy(X), torch.from_numpy(Y)
# use training domain validation `IN SEARCH OF LOST DOMAIN GENERALIZATION`
if self.mode == 'train':
idx = index[(index['date'] <= self.train_maxdate) & (index[domain_name].isin(train_doms))]['index'].to_numpy()
self.index = torch.from_numpy(idx)
elif self.mode == 'validate':
idx = index[(index['date'] <= self.validate_maxdate) & (index['date'] > self.train_maxdate) & (index[domain_name].isin(train_doms))]['index'].to_numpy()
self.index = torch.from_numpy(idx)
elif self.mode == 'test':
idx = index[(index['date'] > self.validate_maxdate) & (index[domain_name].isin(test_doms))]['index'].to_numpy()
self.index = torch.from_numpy(idx)
self.d_emb = 0 # number of embedding categoricals in input
self.d_cov = 2
self.d_lag = 1
self.x_dim = 3
self.y_dim = 1
self.d_dim = num_doms # number of domains
self.x_bin_dim = 0
self.x_con_dim = 3
return X, Y