-
Notifications
You must be signed in to change notification settings - Fork 18
/
checkpointer.py
277 lines (232 loc) · 9.62 KB
/
checkpointer.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
"""Emmental checkpointer."""
import glob
import logging
import os
from shutil import copyfile
from typing import Any, Dict, List, Set, Union
import torch
from torch.optim.lr_scheduler import _LRScheduler
from torch.optim.optimizer import Optimizer
from emmental.meta import Meta
from emmental.model import EmmentalModel
logger = logging.getLogger(__name__)
class Checkpointer(object):
"""Checkpointing class to log train information."""
def __init__(self) -> None:
"""Initialize the checkpointer."""
# Set up checkpoint directory
self.checkpoint_path = Meta.config["logging_config"]["checkpointer_config"][
"checkpoint_path"
]
if self.checkpoint_path is None:
self.checkpoint_path = Meta.log_path
# Create checkpoint directory if necessary
if not os.path.exists(self.checkpoint_path):
os.makedirs(self.checkpoint_path)
# Set up checkpoint frequency
self.checkpoint_freq = (
Meta.config["logging_config"]["evaluation_freq"]
* Meta.config["logging_config"]["checkpointer_config"]["checkpoint_freq"]
)
if self.checkpoint_freq <= 0:
raise ValueError(
f"Invalid checkpoint freq {self.checkpoint_freq}, "
f"must be greater 0."
)
# Set up checkpoint unit
self.checkpoint_unit = Meta.config["logging_config"]["counter_unit"]
logger.info(
f"Save checkpoints at {self.checkpoint_path} every "
f"{self.checkpoint_freq} {self.checkpoint_unit}"
)
# Set up checkpoint metric
self.checkpoint_metric = Meta.config["logging_config"]["checkpointer_config"][
"checkpoint_metric"
]
self.checkpoint_all_metrics = Meta.config["logging_config"][
"checkpointer_config"
]["checkpoint_task_metrics"]
# Collect all metrics to checkpoint
if self.checkpoint_all_metrics is None:
self.checkpoint_all_metrics = dict()
if self.checkpoint_metric:
self.checkpoint_all_metrics.update(self.checkpoint_metric)
# Check evaluation metric mode
for metric, mode in self.checkpoint_all_metrics.items():
if mode not in ["min", "max"]:
raise ValueError(
f"Unrecognized checkpoint metric mode {mode} for metric {metric}, "
f"must be 'min' or 'max'."
)
self.checkpoint_runway = Meta.config["logging_config"]["checkpointer_config"][
"checkpoint_runway"
]
logger.info(
f"No checkpoints saved before {self.checkpoint_runway} "
f"{self.checkpoint_unit}."
)
self.checkpoint_all = Meta.config["logging_config"]["checkpointer_config"][
"checkpoint_all"
]
logger.info(f"Checkpointing all checkpoints: {self.checkpoint_all}.")
self.checkpoint_paths: List[str] = []
# Set up checkpoint clear
self.clear_intermediate_checkpoints = Meta.config["logging_config"][
"checkpointer_config"
]["clear_intermediate_checkpoints"]
self.clear_all_checkpoints = Meta.config["logging_config"][
"checkpointer_config"
]["clear_all_checkpoints"]
# Set up checkpoint flag
self.checkpoint_condition_met = False
self.best_metric_dict: Dict[str, float] = dict()
def checkpoint(
self,
iteration: Union[float, int],
model: EmmentalModel,
optimizer: Optimizer,
lr_scheduler: _LRScheduler,
metric_dict: Dict[str, float],
) -> None:
"""Checkpointing the checkpoint.
Args:
iteration: The current iteration.
model: The model to checkpoint.
optimizer: The optimizer used during training process.
lr_scheduler: Learning rate scheduler.
metric_dict: The metric dict.
"""
# Check the checkpoint_runway condition is met
if iteration < self.checkpoint_runway:
return
elif not self.checkpoint_condition_met and iteration >= self.checkpoint_runway:
self.checkpoint_condition_met = True
logger.info("checkpoint_runway condition has been met. Start checkpoining.")
state_dict = self.collect_state_dict(
iteration, model, optimizer, lr_scheduler, metric_dict
)
checkpoint_path = f"{self.checkpoint_path}/checkpoint_{iteration}.pth"
torch.save(state_dict, checkpoint_path)
logger.info(
f"Save checkpoint of {iteration} {self.checkpoint_unit} "
f"at {checkpoint_path}."
)
if self.checkpoint_all is False:
for path in self.checkpoint_paths:
if os.path.exists(path):
os.remove(path)
self.checkpoint_paths.append(checkpoint_path)
if not set(self.checkpoint_all_metrics.keys()).isdisjoint(
set(metric_dict.keys())
):
new_best_metrics = self.is_new_best(metric_dict)
for metric in new_best_metrics:
copyfile(
checkpoint_path,
f"{self.checkpoint_path}/best_model_"
f"{metric.replace('/', '_')}.pth",
)
logger.info(
f"Save best model of metric {metric} at {self.checkpoint_path}"
f"/best_model_{metric.replace('/', '_')}.pth"
)
def is_new_best(self, metric_dict: Dict[str, float]) -> Set[str]:
"""Update the best score.
Args:
metric_dict: The current metric dict.
Returns:
The updated best metric set.
"""
best_metric = set()
for metric in metric_dict:
if metric not in self.checkpoint_all_metrics:
continue
if metric not in self.best_metric_dict:
self.best_metric_dict[metric] = metric_dict[metric]
best_metric.add(metric)
elif (
self.checkpoint_all_metrics[metric] == "max"
and metric_dict[metric] > self.best_metric_dict[metric]
):
self.best_metric_dict[metric] = metric_dict[metric]
best_metric.add(metric)
elif (
self.checkpoint_all_metrics[metric] == "min"
and metric_dict[metric] < self.best_metric_dict[metric]
):
self.best_metric_dict[metric] = metric_dict[metric]
best_metric.add(metric)
return best_metric
def clear(self) -> None:
"""Clear checkpoints."""
if self.clear_all_checkpoints:
logger.info("Clear all checkpoints.")
file_list = glob.glob(f"{self.checkpoint_path}/*.pth")
for file in file_list:
os.remove(file)
elif self.clear_intermediate_checkpoints:
logger.info("Clear all intermediate checkpoints.")
file_list = glob.glob(f"{self.checkpoint_path}/checkpoint_*.pth")
for file in file_list:
os.remove(file)
def collect_state_dict(
self,
iteration: Union[float, int],
model: EmmentalModel,
optimizer: Optimizer,
lr_scheduler: _LRScheduler,
metric_dict: Dict[str, float],
) -> Dict[str, Any]:
"""Collect the state dict of the model.
Args:
iteration: The current iteration.
model: The model to checkpoint.
optimizer: The optimizer used during training process.
lr_scheduler: Learning rate scheduler.
metric_dict: the metric dict.
Returns:
The state dict.
"""
model_params = {
"name": model.name,
"module_pool": model.collect_state_dict(),
# "task_names": model.task_names,
# "task_flows": model.task_flows,
# "loss_funcs": model.loss_funcs,
# "output_funcs": model.output_funcs,
# "scorers": model.scorers,
}
state_dict = {
"iteration": iteration,
"model": model_params,
"optimizer": optimizer.state_dict(),
"lr_scheduler": lr_scheduler.state_dict() if lr_scheduler else None,
"metric_dict": metric_dict,
}
return state_dict
def load_best_model(self, model: EmmentalModel) -> EmmentalModel:
"""Load the best model from the checkpoint.
Args:
model: The current model.
Returns:
The best model load from the checkpoint.
"""
if list(self.checkpoint_metric.keys())[0] not in self.best_metric_dict:
logger.info("No best model found, use the original model.")
else:
# Load the best model of checkpoint_metric
metric = list(self.checkpoint_metric.keys())[0]
best_model_path = (
f"{self.checkpoint_path}/best_model_{metric.replace('/', '_')}.pth"
)
logger.info(f"Loading the best model from {best_model_path}.")
checkpoint = torch.load(best_model_path, map_location=torch.device("cpu"))
model.name = checkpoint["model"]["name"]
model.load_state_dict(checkpoint["model"]["module_pool"])
# model.task_names = checkpoint["model"]["task_names"]
# model.task_flows = checkpoint["model"]["task_flows"]
# model.loss_funcs = checkpoint["model"]["loss_funcs"]
# model.output_funcs = checkpoint["model"]["output_funcs"]
# model.scorers = checkpoint["model"]["scorers"]
model._move_to_device()
return model