-
Notifications
You must be signed in to change notification settings - Fork 400
/
callback.py
441 lines (351 loc) · 14.1 KB
/
callback.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
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
# Copyright 2022 MosaicML Composer authors
# SPDX-License-Identifier: Apache-2.0
"""Base module for callbacks."""
from __future__ import annotations
import abc
from typing import TYPE_CHECKING, Any
from composer.core.serializable import Serializable
if TYPE_CHECKING:
from composer import Event, State
from composer.loggers import Logger
__all__ = ['Callback']
class Callback(Serializable, abc.ABC):
"""Base class for callbacks.
Callbacks provide hooks that can run at each training loop :class:`.Event`. A callback is similar to
an :class:`.Algorithm` in that they are run on specific events, but it differs from an :class:`.Algorithm`
in that it should not modify the training of the model. By convention, callbacks should not modify the
:class:`.State`. They are typically used to for non-essential recording functions such as logging or timing.
Callbacks can be implemented in two ways:
#. Override the individual methods named for each :class:`.Event`.
For example,
.. doctest::
>>> class MyCallback(Callback):
... def epoch_start(self, state: State, logger: Logger):
... print(f'Epoch: {int(state.timestamp.epoch)}')
>>> # construct trainer object with your callback
>>> trainer = Trainer(
... model=model,
... train_dataloader=train_dataloader,
... eval_dataloader=eval_dataloader,
... optimizers=optimizer,
... max_duration="1ep",
... callbacks=[MyCallback()],
... )
>>> # trainer will run MyCallback whenever the EPOCH_START
>>> # is triggered, like this:
>>> _ = trainer.engine.run_event(Event.EPOCH_START)
Epoch: 0
#. Override :meth:`run_event` if you want a single method to handle all events. If this method is overridden, then
the individual methods corresponding to each event name (such as :meth:`epoch_start`) will no longer be
automatically invoked. For example, if you override :meth:`run_event`, then :meth:`epoch_start` will not be called
on the :attr:`.Event.EPOCH_START` event, :meth:`batch_start` will not be called on the
:attr:`.Event.BATCH_START`, etc. However, you can invoke :meth:`epoch_start`, :meth:`batch_start`, etc. in your
overriding implementation of :meth:`run_event`.
For example,
.. doctest::
>>> class MyCallback(Callback):
... def run_event(self, event: Event, state: State, logger: Logger):
... if event == Event.EPOCH_START:
... print(f'Epoch: {int(state.timestamp.epoch)}')
>>> # construct trainer object with your callback
>>> trainer = Trainer(
... model=model,
... train_dataloader=train_dataloader,
... eval_dataloader=eval_dataloader,
... optimizers=optimizer,
... max_duration="1ep",
... callbacks=[MyCallback()],
... )
>>> # trainer will run MyCallback whenever the EPOCH_START
>>> # is triggered, like this:
>>> _ = trainer.engine.run_event(Event.EPOCH_START)
Epoch: 0
"""
def __init__(self, *args: Any, **kwargs: Any) -> None:
# Stub signature for pyright
del args, kwargs # unused
pass
def run_event(self, event: Event, state: State, logger: Logger) -> None:
"""Called by the engine on each event.
Args:
event (Event): The event.
state (State): The state.
logger (Logger): The logger.
"""
event_cb = getattr(self, event.value)
return event_cb(state, logger)
def init(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.INIT` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def fit_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.FIT_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def epoch_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EPOCH_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def batch_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BATCH_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def after_dataloader(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.AFTER_DATALOADER` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def before_train_batch(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BEFORE_TRAIN_BATCH` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def before_forward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BEFORE_FORWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def after_forward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.AFTER_FORWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def before_loss(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BEFORE_LOSS` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def after_loss(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.AFTER_LOSS` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def before_backward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BEFORE_BACKWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def after_backward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.AFTER_BACKWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def after_train_batch(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.AFTER_TRAIN_BATCH` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def batch_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BATCH_END` event.
.. note::
The following :attr:`.State.timestamp` member variables are
incremented immediately before the :attr:`.Event.BATCH_END` event.
+------------------------------------+
| :attr:`.Timestamp.batch` |
+------------------------------------+
| :attr:`.Timestamp.batch_in_epoch` |
+------------------------------------+
| :attr:`.Timestamp.sample` |
+------------------------------------+
| :attr:`.Timestamp.sample_in_epoch` |
+------------------------------------+
| :attr:`.Timestamp.token` |
+------------------------------------+
| :attr:`.Timestamp.token_in_epoch` |
+------------------------------------+
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def batch_checkpoint(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.BATCH_CHECKPOINT` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def epoch_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EPOCH_END` event.
.. note::
:attr:`.State.timestamp` member variable :attr:`.Timestamp.epoch`
is incremented immediately before :attr:`.Event.EPOCH_END`.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def epoch_checkpoint(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EPOCH_CHECKPOINT` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def predict_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.PREDICT_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def predict_batch_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.PREDICT_BATCH_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def predict_before_forward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.PREDICT_BATCH_FORWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def predict_after_forward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.PREDICT_AFTER_FORWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def predict_batch_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.PREDICT_BATCH_END` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def predict_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.PREDICT_END` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def eval_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EVAL_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def eval_batch_start(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EVAL_BATCH_START` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def eval_before_forward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EVAL_BATCH_FORWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def eval_after_forward(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EVAL_AFTER_FORWARD` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def eval_batch_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EVAL_BATCH_END` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def eval_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.EVAL_END` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def fit_end(self, state: State, logger: Logger) -> None:
"""Called on the :attr:`.Event.FIT_END` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
del state, logger # unused
pass
def close(self, state: State, logger: Logger) -> None:
"""Called whenever the trainer finishes training, even when there is an exception.
It should be used for clean up tasks such as flushing I/O streams and/or
closing any files that may have been opened during the :attr:`.Event.INIT` event.
Args:
state (State): The training state.
logger (Logger): The logger.
"""
pass
def post_close(self) -> None:
"""Called after :meth:`close` has been invoked for each callback.
Very few callbacks should need to implement :meth:`post_close`.
This callback can be used to back up any data that may have
been written by other callbacks during :meth:`close`.
"""
pass