Spaces:
Running
on
Zero
Running
on
Zero
File size: 19,196 Bytes
28c256d |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 |
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
import copy
import logging
from collections import OrderedDict
from typing import TYPE_CHECKING, Any, Optional, Union
import numpy as np
from mmengine.utils import ManagerMixin
from .history_buffer import HistoryBuffer
from .logger import print_log
if TYPE_CHECKING:
import torch
class MessageHub(ManagerMixin):
"""Message hub for component interaction. MessageHub is created and
accessed in the same way as ManagerMixin.
``MessageHub`` will record log information and runtime information. The
log information refers to the learning rate, loss, etc. of the model
during training phase, which will be stored as ``HistoryBuffer``. The
runtime information refers to the iter times, meta information of
runner etc., which will be overwritten by next update.
Args:
name (str): Name of message hub used to get corresponding instance
globally.
log_scalars (dict, optional): Each key-value pair in the
dictionary is the name of the log information such as "loss", "lr",
"metric" and their corresponding values. The type of value must be
HistoryBuffer. Defaults to None.
runtime_info (dict, optional): Each key-value pair in the
dictionary is the name of the runtime information and their
corresponding values. Defaults to None.
resumed_keys (dict, optional): Each key-value pair in the
dictionary decides whether the key in :attr:`_log_scalars` and
:attr:`_runtime_info` will be serialized.
Note:
Key in :attr:`_resumed_keys` belongs to :attr:`_log_scalars` or
:attr:`_runtime_info`. The corresponding value cannot be set
repeatedly.
Examples:
>>> # create empty `MessageHub`.
>>> message_hub1 = MessageHub('name')
>>> log_scalars = dict(loss=HistoryBuffer())
>>> runtime_info = dict(task='task')
>>> resumed_keys = dict(loss=True)
>>> # create `MessageHub` from data.
>>> message_hub2 = MessageHub(
>>> name='name',
>>> log_scalars=log_scalars,
>>> runtime_info=runtime_info,
>>> resumed_keys=resumed_keys)
"""
def __init__(self,
name: str,
log_scalars: Optional[dict] = None,
runtime_info: Optional[dict] = None,
resumed_keys: Optional[dict] = None):
super().__init__(name)
self._log_scalars = self._parse_input('log_scalars', log_scalars)
self._runtime_info = self._parse_input('runtime_info', runtime_info)
self._resumed_keys = self._parse_input('resumed_keys', resumed_keys)
for value in self._log_scalars.values():
assert isinstance(value, HistoryBuffer), \
("The type of log_scalars'value must be HistoryBuffer, but "
f'got {type(value)}')
for key in self._resumed_keys.keys():
assert key in self._log_scalars or key in self._runtime_info, \
('Key in `resumed_keys` must contained in `log_scalars` or '
f'`runtime_info`, but got {key}')
@classmethod
def get_current_instance(cls) -> 'MessageHub':
"""Get latest created ``MessageHub`` instance.
:obj:`MessageHub` can call :meth:`get_current_instance` before any
instance has been created, and return a message hub with the instance
name "mmengine".
Returns:
MessageHub: Empty ``MessageHub`` instance.
"""
if not cls._instance_dict:
cls.get_instance('mmengine')
return super().get_current_instance()
def update_scalar(self,
key: str,
value: Union[int, float, np.ndarray, 'torch.Tensor'],
count: int = 1,
resumed: bool = True) -> None:
"""Update :attr:_log_scalars.
Update ``HistoryBuffer`` in :attr:`_log_scalars`. If corresponding key
``HistoryBuffer`` has been created, ``value`` and ``count`` is the
argument of ``HistoryBuffer.update``, Otherwise, ``update_scalar``
will create an ``HistoryBuffer`` with value and count via the
constructor of ``HistoryBuffer``.
Examples:
>>> message_hub = MessageHub(name='name')
>>> # create loss `HistoryBuffer` with value=1, count=1
>>> message_hub.update_scalar('loss', 1)
>>> # update loss `HistoryBuffer` with value
>>> message_hub.update_scalar('loss', 3)
>>> message_hub.update_scalar('loss', 3, resumed=False)
AssertionError: loss used to be true, but got false now. resumed
keys cannot be modified repeatedly'
Note:
The ``resumed`` argument needs to be consistent for the same
``key``.
Args:
key (str): Key of ``HistoryBuffer``.
value (torch.Tensor or np.ndarray or int or float): Value of log.
count (torch.Tensor or np.ndarray or int or float): Accumulation
times of log, defaults to 1. `count` will be used in smooth
statistics.
resumed (str): Whether the corresponding ``HistoryBuffer``
could be resumed. Defaults to True.
"""
self._set_resumed_keys(key, resumed)
checked_value = self._get_valid_value(value)
assert isinstance(count, int), (
f'The type of count must be int. but got {type(count): {count}}')
if key in self._log_scalars:
self._log_scalars[key].update(checked_value, count)
else:
self._log_scalars[key] = HistoryBuffer([checked_value], [count])
def update_scalars(self, log_dict: dict, resumed: bool = True) -> None:
"""Update :attr:`_log_scalars` with a dict.
``update_scalars`` iterates through each pair of log_dict key-value,
and calls ``update_scalar``. If type of value is dict, the value should
be ``dict(value=xxx) or dict(value=xxx, count=xxx)``. Item in
``log_dict`` has the same resume option.
Note:
The ``resumed`` argument needs to be consistent for the same
``log_dict``.
Args:
log_dict (str): Used for batch updating :attr:`_log_scalars`.
resumed (bool): Whether all ``HistoryBuffer`` referred in
log_dict should be resumed. Defaults to True.
Examples:
>>> message_hub = MessageHub.get_instance('mmengine')
>>> log_dict = dict(a=1, b=2, c=3)
>>> message_hub.update_scalars(log_dict)
>>> # The default count of `a`, `b` and `c` is 1.
>>> log_dict = dict(a=1, b=2, c=dict(value=1, count=2))
>>> message_hub.update_scalars(log_dict)
>>> # The count of `c` is 2.
"""
assert isinstance(log_dict, dict), ('`log_dict` must be a dict!, '
f'but got {type(log_dict)}')
for log_name, log_val in log_dict.items():
if isinstance(log_val, dict):
assert 'value' in log_val, \
f'value must be defined in {log_val}'
count = self._get_valid_value(log_val.get('count', 1))
value = log_val['value']
else:
count = 1
value = log_val
assert isinstance(count,
int), ('The type of count must be int. but got '
f'{type(count): {count}}')
self.update_scalar(log_name, value, count, resumed)
def update_info(self, key: str, value: Any, resumed: bool = True) -> None:
"""Update runtime information.
The key corresponding runtime information will be overwritten each
time calling ``update_info``.
Note:
The ``resumed`` argument needs to be consistent for the same
``key``.
Examples:
>>> message_hub = MessageHub(name='name')
>>> message_hub.update_info('iter', 100)
Args:
key (str): Key of runtime information.
value (Any): Value of runtime information.
resumed (bool): Whether the corresponding ``HistoryBuffer``
could be resumed.
"""
self._set_resumed_keys(key, resumed)
self._runtime_info[key] = value
def pop_info(self, key: str, default: Optional[Any] = None) -> Any:
"""Remove runtime information by key. If the key does not exist, this
method will return the default value.
Args:
key (str): Key of runtime information.
default (Any, optional): The default returned value for the
given key.
Returns:
Any: The runtime information if the key exists.
"""
return self._runtime_info.pop(key, default)
def update_info_dict(self, info_dict: dict, resumed: bool = True) -> None:
"""Update runtime information with dictionary.
The key corresponding runtime information will be overwritten each
time calling ``update_info``.
Note:
The ``resumed`` argument needs to be consistent for the same
``info_dict``.
Examples:
>>> message_hub = MessageHub(name='name')
>>> message_hub.update_info({'iter': 100})
Args:
info_dict (str): Runtime information dictionary.
resumed (bool): Whether the corresponding ``HistoryBuffer``
could be resumed.
"""
assert isinstance(info_dict, dict), ('`log_dict` must be a dict!, '
f'but got {type(info_dict)}')
for key, value in info_dict.items():
self.update_info(key, value, resumed=resumed)
def _set_resumed_keys(self, key: str, resumed: bool) -> None:
"""Set corresponding resumed keys.
This method is called by ``update_scalar``, ``update_scalars`` and
``update_info`` to set the corresponding key is true or false in
:attr:`_resumed_keys`.
Args:
key (str): Key of :attr:`_log_scalrs` or :attr:`_runtime_info`.
resumed (bool): Whether the corresponding ``HistoryBuffer``
could be resumed.
"""
if key not in self._resumed_keys:
self._resumed_keys[key] = resumed
else:
assert self._resumed_keys[key] == resumed, \
f'{key} used to be {self._resumed_keys[key]}, but got ' \
'{resumed} now. resumed keys cannot be modified repeatedly.'
@property
def log_scalars(self) -> OrderedDict:
"""Get all ``HistoryBuffer`` instances.
Note:
Considering the large memory footprint of history buffers in the
post-training, :meth:`get_scalar` will return a reference of
history buffer rather than a copy.
Returns:
OrderedDict: All ``HistoryBuffer`` instances.
"""
return self._log_scalars
@property
def runtime_info(self) -> OrderedDict:
"""Get all runtime information.
Returns:
OrderedDict: A copy of all runtime information.
"""
return self._runtime_info
def get_scalar(self, key: str) -> HistoryBuffer:
"""Get ``HistoryBuffer`` instance by key.
Note:
Considering the large memory footprint of history buffers in the
post-training, :meth:`get_scalar` will not return a reference of
history buffer rather than a copy.
Args:
key (str): Key of ``HistoryBuffer``.
Returns:
HistoryBuffer: Corresponding ``HistoryBuffer`` instance if the
key exists.
"""
if key not in self.log_scalars:
raise KeyError(f'{key} is not found in Messagehub.log_buffers: '
f'instance name is: {MessageHub.instance_name}')
return self.log_scalars[key]
def get_info(self, key: str, default: Optional[Any] = None) -> Any:
"""Get runtime information by key. If the key does not exist, this
method will return default information.
Args:
key (str): Key of runtime information.
default (Any, optional): The default returned value for the
given key.
Returns:
Any: A copy of corresponding runtime information if the key exists.
"""
if key not in self.runtime_info:
return default
else:
# TODO: There are restrictions on objects that can be saved
# return copy.deepcopy(self._runtime_info[key])
return self._runtime_info[key]
def _get_valid_value(
self,
value: Union['torch.Tensor', np.ndarray, np.number, int, float],
) -> Union[int, float]:
"""Convert value to python built-in type.
Args:
value (torch.Tensor or np.ndarray or np.number or int or float):
value of log.
Returns:
float or int: python built-in type value.
"""
if isinstance(value, (np.ndarray, np.number)):
assert value.size == 1
value = value.item()
elif isinstance(value, (int, float)):
value = value
else:
# check whether value is torch.Tensor but don't want
# to import torch in this file
assert hasattr(value, 'numel') and value.numel() == 1
value = value.item()
return value # type: ignore
def state_dict(self) -> dict:
"""Returns a dictionary containing log scalars, runtime information and
resumed keys, which should be resumed.
The returned ``state_dict`` can be loaded by :meth:`load_state_dict`.
Returns:
dict: A dictionary contains ``log_scalars``, ``runtime_info`` and
``resumed_keys``.
"""
saved_scalars = OrderedDict()
saved_info = OrderedDict()
for key, value in self._log_scalars.items():
if self._resumed_keys.get(key, False):
saved_scalars[key] = copy.deepcopy(value)
for key, value in self._runtime_info.items():
if self._resumed_keys.get(key, False):
try:
saved_info[key] = copy.deepcopy(value)
except: # noqa: E722
print_log(
f'{key} in message_hub cannot be copied, '
f'just return its reference. ',
logger='current',
level=logging.WARNING)
saved_info[key] = value
return dict(
log_scalars=saved_scalars,
runtime_info=saved_info,
resumed_keys=self._resumed_keys)
def load_state_dict(self, state_dict: Union['MessageHub', dict]) -> None:
"""Loads log scalars, runtime information and resumed keys from
``state_dict`` or ``message_hub``.
If ``state_dict`` is a dictionary returned by :meth:`state_dict`, it
will only make copies of data which should be resumed from the source
``message_hub``.
If ``state_dict`` is a ``message_hub`` instance, it will make copies of
all data from the source message_hub. We suggest to load data from
``dict`` rather than a ``MessageHub`` instance.
Args:
state_dict (dict or MessageHub): A dictionary contains key
``log_scalars`` ``runtime_info`` and ``resumed_keys``, or a
MessageHub instance.
"""
if isinstance(state_dict, dict):
for key in ('log_scalars', 'runtime_info', 'resumed_keys'):
assert key in state_dict, (
'The loaded `state_dict` of `MessageHub` must contain '
f'key: `{key}`')
# The old `MessageHub` could save non-HistoryBuffer `log_scalars`,
# therefore the loaded `log_scalars` needs to be filtered.
for key, value in state_dict['log_scalars'].items():
if not isinstance(value, HistoryBuffer):
print_log(
f'{key} in message_hub is not HistoryBuffer, '
f'just skip resuming it.',
logger='current',
level=logging.WARNING)
continue
self.log_scalars[key] = value
for key, value in state_dict['runtime_info'].items():
try:
self._runtime_info[key] = copy.deepcopy(value)
except: # noqa: E722
print_log(
f'{key} in message_hub cannot be copied, '
f'just return its reference.',
logger='current',
level=logging.WARNING)
self._runtime_info[key] = value
for key, value in state_dict['resumed_keys'].items():
if key not in set(self.log_scalars.keys()) | \
set(self._runtime_info.keys()):
print_log(
f'resumed key: {key} is not defined in message_hub, '
f'just skip resuming this key.',
logger='current',
level=logging.WARNING)
continue
elif not value:
print_log(
f'Although resumed key: {key} is False, {key} '
'will still be loaded this time. This key will '
'not be saved by the next calling of '
'`MessageHub.state_dict()`',
logger='current',
level=logging.WARNING)
self._resumed_keys[key] = value
# Since some checkpoints saved serialized `message_hub` instance,
# `load_state_dict` support loading `message_hub` instance for
# compatibility
else:
self._log_scalars = copy.deepcopy(state_dict._log_scalars)
self._runtime_info = copy.deepcopy(state_dict._runtime_info)
self._resumed_keys = copy.deepcopy(state_dict._resumed_keys)
def _parse_input(self, name: str, value: Any) -> OrderedDict:
"""Parse input value.
Args:
name (str): name of input value.
value (Any): Input value.
Returns:
dict: Parsed input value.
"""
if value is None:
return OrderedDict()
elif isinstance(value, dict):
return OrderedDict(value)
else:
raise TypeError(f'{name} should be a dict or `None`, but '
f'got {type(name)}')
|