-
Notifications
You must be signed in to change notification settings - Fork 189
/
estimator.py
182 lines (157 loc) · 5.61 KB
/
estimator.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
from typing import NamedTuple, Optional
from functools import partial
import numpy as np
import torch
import torch.nn as nn
from torch.utils import data
from torch.utils.data import DataLoader
from gluonts.core.component import validated
from gluonts.dataset.common import Dataset
from gluonts.model.estimator import Estimator
from gluonts.torch.model.predictor import PyTorchPredictor
from gluonts.transform import SelectFields, Transformation
from pts import Trainer
from pts.model import get_module_forward_input_names
from pts.dataset.loader import TransformedIterableDataset
class TrainOutput(NamedTuple):
transformation: Transformation
trained_net: nn.Module
predictor: PyTorchPredictor
class PyTorchEstimator(Estimator):
@validated()
def __init__(
self, trainer: Trainer, lead_time: int = 0, dtype: np.dtype = np.float32
) -> None:
super().__init__(lead_time=lead_time)
self.trainer = trainer
self.dtype = dtype
def create_transformation(self) -> Transformation:
"""
Create and return the transformation needed for training and inference.
Returns
-------
Transformation
The transformation that will be applied entry-wise to datasets,
at training and inference time.
"""
raise NotImplementedError
def create_instance_splitter(self, mode: str) -> Transformation:
"""
Create and return the instance splitter needed for training, validation or testing.
Returns
-------
Transformation
The InstanceSplitter that will be applied entry-wise to datasets,
at training, validation and inference time based on mode.
"""
raise NotImplementedError
def create_training_network(self, device: torch.device) -> nn.Module:
"""
Create and return the network used for training (i.e., computing the
loss).
Returns
-------
nn.Module
The network that computes the loss given input data.
"""
raise NotImplementedError
def create_predictor(
self,
transformation: Transformation,
trained_network: nn.Module,
device: torch.device,
) -> PyTorchPredictor:
"""
Create and return a predictor object.
Returns
-------
Predictor
A predictor wrapping a `nn.Module` used for inference.
"""
raise NotImplementedError
def train_model(
self,
training_data: Dataset,
validation_data: Optional[Dataset] = None,
num_workers: int = 0,
prefetch_factor: int = 2,
shuffle_buffer_length: Optional[int] = None,
cache_data: bool = False,
**kwargs,
) -> TrainOutput:
transformation = self.create_transformation()
trained_net = self.create_training_network(self.trainer.device)
input_names = get_module_forward_input_names(trained_net)
training_instance_splitter = self.create_instance_splitter("training")
training_iter_dataset = TransformedIterableDataset(
dataset=training_data,
transform=transformation
+ training_instance_splitter
+ SelectFields(input_names),
is_train=True,
shuffle_buffer_length=shuffle_buffer_length,
cache_data=cache_data,
)
training_data_loader = DataLoader(
training_iter_dataset,
batch_size=self.trainer.batch_size,
num_workers=num_workers,
prefetch_factor=prefetch_factor,
pin_memory=True,
worker_init_fn=self._worker_init_fn,
**kwargs,
)
validation_data_loader = None
if validation_data is not None:
validation_instance_splitter = self.create_instance_splitter("validation")
validation_iter_dataset = TransformedIterableDataset(
dataset=validation_data,
transform=transformation
+ validation_instance_splitter
+ SelectFields(input_names),
is_train=True,
cache_data=cache_data,
)
validation_data_loader = DataLoader(
validation_iter_dataset,
batch_size=self.trainer.batch_size,
num_workers=num_workers,
prefetch_factor=prefetch_factor,
pin_memory=True,
worker_init_fn=self._worker_init_fn,
**kwargs,
)
self.trainer(
net=trained_net,
train_iter=training_data_loader,
validation_iter=validation_data_loader,
)
return TrainOutput(
transformation=transformation,
trained_net=trained_net,
predictor=self.create_predictor(
transformation, trained_net, self.trainer.device
),
)
@staticmethod
def _worker_init_fn(worker_id):
np.random.seed(np.random.get_state()[1][0] + worker_id)
def train(
self,
training_data: Dataset,
validation_data: Optional[Dataset] = None,
num_workers: int = 0,
prefetch_factor: int = 2,
shuffle_buffer_length: Optional[int] = None,
cache_data: bool = False,
**kwargs,
) -> PyTorchPredictor:
return self.train_model(
training_data,
validation_data,
num_workers=num_workers,
prefetch_factor=prefetch_factor,
shuffle_buffer_length=shuffle_buffer_length,
cache_data=cache_data,
**kwargs,
).predictor