File size: 16,024 Bytes
18dd6ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# -*- coding: utf-8 -*-
# Copyright (c) Facebook, Inc. and its affiliates.

import logging
import numpy as np
import time
import weakref
from typing import List, Mapping, Optional
import torch
from torch.nn.parallel import DataParallel, DistributedDataParallel

import annotator.oneformer.detectron2.utils.comm as comm
from annotator.oneformer.detectron2.utils.events import EventStorage, get_event_storage
from annotator.oneformer.detectron2.utils.logger import _log_api_usage

__all__ = ["HookBase", "TrainerBase", "SimpleTrainer", "AMPTrainer"]


class HookBase:
    """
    Base class for hooks that can be registered with :class:`TrainerBase`.

    Each hook can implement 4 methods. The way they are called is demonstrated
    in the following snippet:
    ::
        hook.before_train()
        for iter in range(start_iter, max_iter):
            hook.before_step()
            trainer.run_step()
            hook.after_step()
        iter += 1
        hook.after_train()

    Notes:
        1. In the hook method, users can access ``self.trainer`` to access more
           properties about the context (e.g., model, current iteration, or config
           if using :class:`DefaultTrainer`).

        2. A hook that does something in :meth:`before_step` can often be
           implemented equivalently in :meth:`after_step`.
           If the hook takes non-trivial time, it is strongly recommended to
           implement the hook in :meth:`after_step` instead of :meth:`before_step`.
           The convention is that :meth:`before_step` should only take negligible time.

           Following this convention will allow hooks that do care about the difference
           between :meth:`before_step` and :meth:`after_step` (e.g., timer) to
           function properly.

    """

    trainer: "TrainerBase" = None
    """
    A weak reference to the trainer object. Set by the trainer when the hook is registered.
    """

    def before_train(self):
        """
        Called before the first iteration.
        """
        pass

    def after_train(self):
        """
        Called after the last iteration.
        """
        pass

    def before_step(self):
        """
        Called before each iteration.
        """
        pass

    def after_backward(self):
        """
        Called after the backward pass of each iteration.
        """
        pass

    def after_step(self):
        """
        Called after each iteration.
        """
        pass

    def state_dict(self):
        """
        Hooks are stateless by default, but can be made checkpointable by
        implementing `state_dict` and `load_state_dict`.
        """
        return {}


class TrainerBase:
    """
    Base class for iterative trainer with hooks.

    The only assumption we made here is: the training runs in a loop.
    A subclass can implement what the loop is.
    We made no assumptions about the existence of dataloader, optimizer, model, etc.

    Attributes:
        iter(int): the current iteration.

        start_iter(int): The iteration to start with.
            By convention the minimum possible value is 0.

        max_iter(int): The iteration to end training.

        storage(EventStorage): An EventStorage that's opened during the course of training.
    """

    def __init__(self) -> None:
        self._hooks: List[HookBase] = []
        self.iter: int = 0
        self.start_iter: int = 0
        self.max_iter: int
        self.storage: EventStorage
        _log_api_usage("trainer." + self.__class__.__name__)

    def register_hooks(self, hooks: List[Optional[HookBase]]) -> None:
        """
        Register hooks to the trainer. The hooks are executed in the order
        they are registered.

        Args:
            hooks (list[Optional[HookBase]]): list of hooks
        """
        hooks = [h for h in hooks if h is not None]
        for h in hooks:
            assert isinstance(h, HookBase)
            # To avoid circular reference, hooks and trainer cannot own each other.
            # This normally does not matter, but will cause memory leak if the
            # involved objects contain __del__:
            # See http://engineering.hearsaysocial.com/2013/06/16/circular-references-in-python/
            h.trainer = weakref.proxy(self)
        self._hooks.extend(hooks)

    def train(self, start_iter: int, max_iter: int):
        """
        Args:
            start_iter, max_iter (int): See docs above
        """
        logger = logging.getLogger(__name__)
        logger.info("Starting training from iteration {}".format(start_iter))

        self.iter = self.start_iter = start_iter
        self.max_iter = max_iter

        with EventStorage(start_iter) as self.storage:
            try:
                self.before_train()
                for self.iter in range(start_iter, max_iter):
                    self.before_step()
                    self.run_step()
                    self.after_step()
                # self.iter == max_iter can be used by `after_train` to
                # tell whether the training successfully finished or failed
                # due to exceptions.
                self.iter += 1
            except Exception:
                logger.exception("Exception during training:")
                raise
            finally:
                self.after_train()

    def before_train(self):
        for h in self._hooks:
            h.before_train()

    def after_train(self):
        self.storage.iter = self.iter
        for h in self._hooks:
            h.after_train()

    def before_step(self):
        # Maintain the invariant that storage.iter == trainer.iter
        # for the entire execution of each step
        self.storage.iter = self.iter

        for h in self._hooks:
            h.before_step()

    def after_backward(self):
        for h in self._hooks:
            h.after_backward()

    def after_step(self):
        for h in self._hooks:
            h.after_step()

    def run_step(self):
        raise NotImplementedError

    def state_dict(self):
        ret = {"iteration": self.iter}
        hooks_state = {}
        for h in self._hooks:
            sd = h.state_dict()
            if sd:
                name = type(h).__qualname__
                if name in hooks_state:
                    # TODO handle repetitive stateful hooks
                    continue
                hooks_state[name] = sd
        if hooks_state:
            ret["hooks"] = hooks_state
        return ret

    def load_state_dict(self, state_dict):
        logger = logging.getLogger(__name__)
        self.iter = state_dict["iteration"]
        for key, value in state_dict.get("hooks", {}).items():
            for h in self._hooks:
                try:
                    name = type(h).__qualname__
                except AttributeError:
                    continue
                if name == key:
                    h.load_state_dict(value)
                    break
            else:
                logger.warning(f"Cannot find the hook '{key}', its state_dict is ignored.")


class SimpleTrainer(TrainerBase):
    """
    A simple trainer for the most common type of task:
    single-cost single-optimizer single-data-source iterative optimization,
    optionally using data-parallelism.
    It assumes that every step, you:

    1. Compute the loss with a data from the data_loader.
    2. Compute the gradients with the above loss.
    3. Update the model with the optimizer.

    All other tasks during training (checkpointing, logging, evaluation, LR schedule)
    are maintained by hooks, which can be registered by :meth:`TrainerBase.register_hooks`.

    If you want to do anything fancier than this,
    either subclass TrainerBase and implement your own `run_step`,
    or write your own training loop.
    """

    def __init__(self, model, data_loader, optimizer, gather_metric_period=1):
        """
        Args:
            model: a torch Module. Takes a data from data_loader and returns a
                dict of losses.
            data_loader: an iterable. Contains data to be used to call model.
            optimizer: a torch optimizer.
            gather_metric_period: an int. Every gather_metric_period iterations
                the metrics are gathered from all the ranks to rank 0 and logged.
        """
        super().__init__()

        """
        We set the model to training mode in the trainer.
        However it's valid to train a model that's in eval mode.
        If you want your model (or a submodule of it) to behave
        like evaluation during training, you can overwrite its train() method.
        """
        model.train()

        self.model = model
        self.data_loader = data_loader
        # to access the data loader iterator, call `self._data_loader_iter`
        self._data_loader_iter_obj = None
        self.optimizer = optimizer
        self.gather_metric_period = gather_metric_period

    def run_step(self):
        """
        Implement the standard training logic described above.
        """
        assert self.model.training, "[SimpleTrainer] model was changed to eval mode!"
        start = time.perf_counter()
        """
        If you want to do something with the data, you can wrap the dataloader.
        """
        data = next(self._data_loader_iter)
        data_time = time.perf_counter() - start

        """
        If you want to do something with the losses, you can wrap the model.
        """
        loss_dict = self.model(data)
        if isinstance(loss_dict, torch.Tensor):
            losses = loss_dict
            loss_dict = {"total_loss": loss_dict}
        else:
            losses = sum(loss_dict.values())

        """
        If you need to accumulate gradients or do something similar, you can
        wrap the optimizer with your custom `zero_grad()` method.
        """
        self.optimizer.zero_grad()
        losses.backward()

        self.after_backward()

        self._write_metrics(loss_dict, data_time)

        """
        If you need gradient clipping/scaling or other processing, you can
        wrap the optimizer with your custom `step()` method. But it is
        suboptimal as explained in https://arxiv.org/abs/2006.15704 Sec 3.2.4
        """
        self.optimizer.step()

    @property
    def _data_loader_iter(self):
        # only create the data loader iterator when it is used
        if self._data_loader_iter_obj is None:
            self._data_loader_iter_obj = iter(self.data_loader)
        return self._data_loader_iter_obj

    def reset_data_loader(self, data_loader_builder):
        """
        Delete and replace the current data loader with a new one, which will be created
        by calling `data_loader_builder` (without argument).
        """
        del self.data_loader
        data_loader = data_loader_builder()
        self.data_loader = data_loader
        self._data_loader_iter_obj = None

    def _write_metrics(
        self,
        loss_dict: Mapping[str, torch.Tensor],
        data_time: float,
        prefix: str = "",
    ) -> None:
        if (self.iter + 1) % self.gather_metric_period == 0:
            SimpleTrainer.write_metrics(loss_dict, data_time, prefix)

    @staticmethod
    def write_metrics(
        loss_dict: Mapping[str, torch.Tensor],
        data_time: float,
        prefix: str = "",
    ) -> None:
        """
        Args:
            loss_dict (dict): dict of scalar losses
            data_time (float): time taken by the dataloader iteration
            prefix (str): prefix for logging keys
        """
        metrics_dict = {k: v.detach().cpu().item() for k, v in loss_dict.items()}
        metrics_dict["data_time"] = data_time

        # Gather metrics among all workers for logging
        # This assumes we do DDP-style training, which is currently the only
        # supported method in detectron2.
        all_metrics_dict = comm.gather(metrics_dict)

        if comm.is_main_process():
            storage = get_event_storage()

            # data_time among workers can have high variance. The actual latency
            # caused by data_time is the maximum among workers.
            data_time = np.max([x.pop("data_time") for x in all_metrics_dict])
            storage.put_scalar("data_time", data_time)

            # average the rest metrics
            metrics_dict = {
                k: np.mean([x[k] for x in all_metrics_dict]) for k in all_metrics_dict[0].keys()
            }
            total_losses_reduced = sum(metrics_dict.values())
            if not np.isfinite(total_losses_reduced):
                raise FloatingPointError(
                    f"Loss became infinite or NaN at iteration={storage.iter}!\n"
                    f"loss_dict = {metrics_dict}"
                )

            storage.put_scalar("{}total_loss".format(prefix), total_losses_reduced)
            if len(metrics_dict) > 1:
                storage.put_scalars(**metrics_dict)

    def state_dict(self):
        ret = super().state_dict()
        ret["optimizer"] = self.optimizer.state_dict()
        return ret

    def load_state_dict(self, state_dict):
        super().load_state_dict(state_dict)
        self.optimizer.load_state_dict(state_dict["optimizer"])


class AMPTrainer(SimpleTrainer):
    """
    Like :class:`SimpleTrainer`, but uses PyTorch's native automatic mixed precision
    in the training loop.
    """

    def __init__(
        self,
        model,
        data_loader,
        optimizer,
        gather_metric_period=1,
        grad_scaler=None,
        precision: torch.dtype = torch.float16,
        log_grad_scaler: bool = False,
    ):
        """
        Args:
            model, data_loader, optimizer, gather_metric_period: same as in :class:`SimpleTrainer`.
            grad_scaler: torch GradScaler to automatically scale gradients.
            precision: torch.dtype as the target precision to cast to in computations
        """
        unsupported = "AMPTrainer does not support single-process multi-device training!"
        if isinstance(model, DistributedDataParallel):
            assert not (model.device_ids and len(model.device_ids) > 1), unsupported
        assert not isinstance(model, DataParallel), unsupported

        super().__init__(model, data_loader, optimizer, gather_metric_period)

        if grad_scaler is None:
            from torch.cuda.amp import GradScaler

            grad_scaler = GradScaler()
        self.grad_scaler = grad_scaler
        self.precision = precision
        self.log_grad_scaler = log_grad_scaler

    def run_step(self):
        """
        Implement the AMP training logic.
        """
        assert self.model.training, "[AMPTrainer] model was changed to eval mode!"
        assert torch.cuda.is_available(), "[AMPTrainer] CUDA is required for AMP training!"
        from torch.cuda.amp import autocast

        start = time.perf_counter()
        data = next(self._data_loader_iter)
        data_time = time.perf_counter() - start

        with autocast(dtype=self.precision):
            loss_dict = self.model(data)
            if isinstance(loss_dict, torch.Tensor):
                losses = loss_dict
                loss_dict = {"total_loss": loss_dict}
            else:
                losses = sum(loss_dict.values())

        self.optimizer.zero_grad()
        self.grad_scaler.scale(losses).backward()

        if self.log_grad_scaler:
            storage = get_event_storage()
            storage.put_scalar("[metric]grad_scaler", self.grad_scaler.get_scale())

        self.after_backward()

        self._write_metrics(loss_dict, data_time)

        self.grad_scaler.step(self.optimizer)
        self.grad_scaler.update()

    def state_dict(self):
        ret = super().state_dict()
        ret["grad_scaler"] = self.grad_scaler.state_dict()
        return ret

    def load_state_dict(self, state_dict):
        super().load_state_dict(state_dict)
        self.grad_scaler.load_state_dict(state_dict["grad_scaler"])