-
-
Notifications
You must be signed in to change notification settings - Fork 84
/
saits.py
411 lines (356 loc) · 13.9 KB
/
saits.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
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
"""
PyTorch SAITS model for the time-series imputation task.
Notes
-----
Partial implementation uses code from https://github.com/WenjieDu/SAITS.
"""
# Created by Wenjie Du <[email protected]>
# License: GPL-v3
from typing import Tuple, Union, Optional
import h5py
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from pypots.data.base import BaseDataset
from pypots.data.dataset_for_mit import DatasetForMIT
from pypots.imputation.base import BaseNNImputer
from pypots.imputation.transformer import EncoderLayer, PositionalEncoding
from pypots.utils.metrics import cal_mae
class _SAITS(nn.Module):
def __init__(
self,
n_layers: int,
d_time: int,
d_feature: int,
d_model: int,
d_inner: int,
n_head: int,
d_k: int,
d_v: int,
dropout: float,
diagonal_attention_mask: bool = True,
ORT_weight: float = 1,
MIT_weight: float = 1,
):
super().__init__()
self.n_layers = n_layers
actual_d_feature = d_feature * 2
self.ORT_weight = ORT_weight
self.MIT_weight = MIT_weight
self.layer_stack_for_first_block = nn.ModuleList(
[
EncoderLayer(
d_time,
actual_d_feature,
d_model,
d_inner,
n_head,
d_k,
d_v,
dropout,
0,
diagonal_attention_mask,
)
for _ in range(n_layers)
]
)
self.layer_stack_for_second_block = nn.ModuleList(
[
EncoderLayer(
d_time,
actual_d_feature,
d_model,
d_inner,
n_head,
d_k,
d_v,
dropout,
0,
diagonal_attention_mask,
)
for _ in range(n_layers)
]
)
self.dropout = nn.Dropout(p=dropout)
self.position_enc = PositionalEncoding(d_model, n_position=d_time)
# for operation on time dim
self.embedding_1 = nn.Linear(actual_d_feature, d_model)
self.reduce_dim_z = nn.Linear(d_model, d_feature)
# for operation on measurement dim
self.embedding_2 = nn.Linear(actual_d_feature, d_model)
self.reduce_dim_beta = nn.Linear(d_model, d_feature)
self.reduce_dim_gamma = nn.Linear(d_feature, d_feature)
# for delta decay factor
self.weight_combine = nn.Linear(d_feature + d_time, d_feature)
def _process(self, inputs: dict) -> Tuple[torch.Tensor, list]:
X, masks = inputs["X"], inputs["missing_mask"]
# first DMSA block
input_X_for_first = torch.cat([X, masks], dim=2)
input_X_for_first = self.embedding_1(input_X_for_first)
enc_output = self.dropout(
self.position_enc(input_X_for_first)
) # namely, term e in the math equation
for encoder_layer in self.layer_stack_for_first_block:
enc_output, _ = encoder_layer(enc_output)
X_tilde_1 = self.reduce_dim_z(enc_output)
X_prime = masks * X + (1 - masks) * X_tilde_1
# second DMSA block
input_X_for_second = torch.cat([X_prime, masks], dim=2)
input_X_for_second = self.embedding_2(input_X_for_second)
enc_output = self.position_enc(
input_X_for_second
) # namely term alpha in math algo
for encoder_layer in self.layer_stack_for_second_block:
enc_output, attn_weights = encoder_layer(enc_output)
X_tilde_2 = self.reduce_dim_gamma(F.relu(self.reduce_dim_beta(enc_output)))
# attention-weighted combine
attn_weights = attn_weights.squeeze(dim=1) # namely term A_hat in Eq.
if len(attn_weights.shape) == 4:
# if having more than 1 head, then average attention weights from all heads
attn_weights = torch.transpose(attn_weights, 1, 3)
attn_weights = attn_weights.mean(dim=3)
attn_weights = torch.transpose(attn_weights, 1, 2)
# namely term eta
combining_weights = torch.sigmoid(
self.weight_combine(torch.cat([masks, attn_weights], dim=2))
)
# combine X_tilde_1 and X_tilde_2
X_tilde_3 = (1 - combining_weights) * X_tilde_2 + combining_weights * X_tilde_1
# replace non-missing part with original data
X_c = masks * X + (1 - masks) * X_tilde_3
return X_c, [X_tilde_1, X_tilde_2, X_tilde_3]
def impute(self, inputs: dict) -> torch.Tensor:
imputed_data, _ = self._process(inputs)
return imputed_data
def forward(self, inputs: dict) -> dict:
X, masks = inputs["X"], inputs["missing_mask"]
ORT_loss = 0
imputed_data, [X_tilde_1, X_tilde_2, X_tilde_3] = self._process(inputs)
ORT_loss += cal_mae(X_tilde_1, X, masks)
ORT_loss += cal_mae(X_tilde_2, X, masks)
ORT_loss += cal_mae(X_tilde_3, X, masks)
ORT_loss /= 3
MIT_loss = cal_mae(X_tilde_3, inputs["X_intact"], inputs["indicating_mask"])
# `loss` is always the item for backward propagating to update the model
loss = self.ORT_weight * ORT_loss + self.MIT_weight * MIT_loss
results = {
"imputed_data": imputed_data,
"ORT_loss": ORT_loss,
"MIT_loss": MIT_loss,
"loss": loss, # will be used for backward propagating to update the model
}
return results
class SAITS(BaseNNImputer):
def __init__(
self,
n_steps: int,
n_features: int,
n_layers: int,
d_model: int,
d_inner: int,
n_head: int,
d_k: int,
d_v: int,
dropout: int or float,
diagonal_attention_mask: bool = True,
ORT_weight: int = 1,
MIT_weight: int = 1,
batch_size: int = 32,
epochs: int = 100,
patience: int = None,
learning_rate: float = 1e-3,
weight_decay: float = 1e-5,
num_workers: int = 0,
device: Optional[Union[str, torch.device]] = None,
tb_file_saving_path: str = None,
):
super().__init__(
batch_size,
epochs,
patience,
learning_rate,
weight_decay,
num_workers,
device,
tb_file_saving_path,
)
self.n_steps = n_steps
self.n_features = n_features
# model hype-parameters
self.n_layers = n_layers
self.d_model = d_model
self.d_inner = d_inner
self.n_head = n_head
self.d_k = d_k
self.d_v = d_v
self.dropout = dropout
self.diagonal_attention_mask = diagonal_attention_mask
self.ORT_weight = ORT_weight
self.MIT_weight = MIT_weight
self.model = _SAITS(
self.n_layers,
self.n_steps,
self.n_features,
self.d_model,
self.d_inner,
self.n_head,
self.d_k,
self.d_v,
self.dropout,
self.diagonal_attention_mask,
self.ORT_weight,
self.MIT_weight,
)
self.model = self.model.to(self.device)
self._print_model_size()
def _assemble_input_for_training(self, data: list) -> dict:
"""Assemble the given data into a dictionary for training input.
Parameters
----------
data : list,
A list containing data fetched from Dataset by Dataloader.
Returns
-------
inputs : dict,
A python dictionary contains the input data for model training.
"""
indices, X_intact, X, missing_mask, indicating_mask = map(
lambda x: x.to(self.device), data
)
inputs = {
"X": X,
"X_intact": X_intact,
"missing_mask": missing_mask,
"indicating_mask": indicating_mask,
}
return inputs
def _assemble_input_for_validating(self, data) -> dict:
"""Assemble the given data into a dictionary for validating input.
Notes
-----
The validating data assembling processing is the same as training data assembling.
Parameters
----------
data : list,
A list containing data fetched from Dataset by Dataloader.
Returns
-------
inputs : dict,
A python dictionary contains the input data for model validating.
"""
indices, X, missing_mask = map(lambda x: x.to(self.device), data)
inputs = {
"X": X,
"missing_mask": missing_mask,
}
return inputs
def _assemble_input_for_testing(self, data) -> dict:
"""Assemble the given data into a dictionary for testing input.
Notes
-----
The testing data assembling processing is the same as training data assembling.
Parameters
----------
data : list,
A list containing data fetched from Dataset by Dataloader.
Returns
-------
inputs : dict,
A python dictionary contains the input data for model testing.
"""
return self._assemble_input_for_validating(data)
def fit(
self,
train_set: Union[dict, str],
val_set: Optional[Union[dict, str]] = None,
file_type: str = "h5py",
) -> None:
"""Train the imputer on the given data.
Parameters
----------
train_set : dict or str,
The dataset for model training, should be a dictionary including the key 'X',
or a path string locating a data file.
If it is a dict, X should be array-like of shape [n_samples, sequence length (time steps), n_features],
which is time-series data for training, can contain missing values.
If it is a path string, the path should point to a data file, e.g. a h5 file, which contains
key-value pairs like a dict, and it has to include the key 'X'.
val_set : dict or str,
The dataset for model validating, should be a dictionary including the key 'X',
or a path string locating a data file.
If it is a dict, X should be array-like of shape [n_samples, sequence length (time steps), n_features],
which is time-series data for validating, can contain missing values.
If it is a path string, the path should point to a data file, e.g. a h5 file, which contains
key-value pairs like a dict, and it has to include the key 'X'.
file_type : str, default = "h5py",
The type of the given file if train_set and val_set are path strings.
"""
training_set = DatasetForMIT(train_set, file_type)
training_loader = DataLoader(
training_set,
batch_size=self.batch_size,
shuffle=True,
num_workers=self.num_workers,
)
if val_set is None:
self._train_model(training_loader)
else:
if isinstance(val_set, str):
with h5py.File(val_set, "r") as hf:
# Here we read the whole validation set from the file to mask a portion for validation.
# In PyPOTS, using a file usually because the data is too big. However, the validation set is
# generally shouldn't be too large. For example, we have 1 billion samples for model training.
# We won't take 20% of them as the validation set because we want as much as possible data for the
# training stage to enhance the model's generalization ability. Therefore, 100,000 representative
# samples will be enough to validate the model.
val_set = {
"X": hf["X"][:],
"X_intact": hf["X_intact"][:],
"indicating_mask": hf["indicating_mask"][:],
}
val_set = BaseDataset(val_set)
val_loader = DataLoader(
val_set,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers,
)
self._train_model(training_loader, val_loader)
self.model.load_state_dict(self.best_model_dict)
self.model.eval() # set the model as eval status to freeze it.
def impute(
self,
X: Union[dict, str],
file_type="h5py",
) -> np.ndarray:
"""Impute missing values in the given data with the trained model.
Parameters
----------
X : array-like or str,
The data samples for testing, should be array-like of shape [n_samples, sequence length (time steps),
n_features], or a path string locating a data file, e.g. h5 file.
file_type : str, default = "h5py",
The type of the given file if X is a path string.
Returns
-------
array-like, shape [n_samples, sequence length (time steps), n_features],
Imputed data.
"""
self.model.eval() # set the model as eval status to freeze it.
test_set = BaseDataset(X, file_type)
test_loader = DataLoader(
test_set,
batch_size=self.batch_size,
shuffle=False,
num_workers=self.num_workers,
)
imputation_collector = []
with torch.no_grad():
for idx, data in enumerate(test_loader):
inputs = self._assemble_input_for_testing(data)
imputed_data = self.model.impute(inputs)
imputation_collector.append(imputed_data)
imputation_collector = torch.cat(imputation_collector)
return imputation_collector.cpu().detach().numpy()