callbacks.py 39.4 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

15
import os
16
import time
17
import numbers
L
LiuChiachi 已提交
18 19 20
import warnings

import numpy as np
21

22 23
import paddle
from paddle.distributed import ParallelEnv
24
from paddle.utils import try_import
25 26 27

from .progressbar import ProgressBar

28
__all__ = [
L
LiuChiachi 已提交
29
    'Callback', 'ProgBarLogger', 'ModelCheckpoint', 'VisualDL', 'LRScheduler',
L
LielinJiang 已提交
30
    'EarlyStopping', 'ReduceLROnPlateau'
31
]
32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52


def config_callbacks(callbacks=None,
                     model=None,
                     batch_size=None,
                     epochs=None,
                     steps=None,
                     log_freq=2,
                     verbose=2,
                     save_freq=1,
                     save_dir=None,
                     metrics=None,
                     mode='train'):
    cbks = callbacks or []
    cbks = cbks if isinstance(cbks, (list, tuple)) else [cbks]
    if not any(isinstance(k, ProgBarLogger) for k in cbks) and verbose:
        cbks = [ProgBarLogger(log_freq, verbose=verbose)] + cbks

    if not any(isinstance(k, ModelCheckpoint) for k in cbks):
        cbks = cbks + [ModelCheckpoint(save_freq, save_dir)]

L
LiuChiachi 已提交
53 54 55
    for k in cbks:
        if isinstance(k, EarlyStopping):
            k.save_dir = save_dir
56 57 58
    if not any(isinstance(k, LRScheduler) for k in cbks):
        cbks = cbks + [LRScheduler()]

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
    cbk_list = CallbackList(cbks)
    cbk_list.set_model(model)
    metrics = metrics or [] if mode != 'test' else []
    params = {
        'batch_size': batch_size,
        'epochs': epochs,
        'steps': steps,
        'verbose': verbose,
        'metrics': metrics,
    }
    cbk_list.set_params(params)
    return cbk_list


class CallbackList(object):
    def __init__(self, callbacks=None):
        # copy
        self.callbacks = [c for c in callbacks]
        self.params = {}
        self.model = None

    def append(self, callback):
        self.callbacks.append(callback)

    def __iter__(self):
        return iter(self.callbacks)

    def set_params(self, params):
        for c in self.callbacks:
            c.set_params(params)

    def set_model(self, model):
        for c in self.callbacks:
            c.set_model(model)

    def _call(self, name, *args):
        for c in self.callbacks:
            func = getattr(c, name)
            func(*args)

    def _check_mode(self, mode):
100 101
        assert mode in ['train', 'eval', 'predict'], \
            'mode should be train, eval or predict'
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

    def on_begin(self, mode, logs=None):
        self._check_mode(mode)
        name = 'on_{}_begin'.format(mode)
        self._call(name, logs)

    def on_end(self, mode, logs=None):
        self._check_mode(mode)
        name = 'on_{}_end'.format(mode)
        self._call(name, logs)

    def on_epoch_begin(self, epoch=None, logs=None):
        self._call('on_epoch_begin', epoch, logs)

    def on_epoch_end(self, epoch=None, logs=None):
        self._call('on_epoch_end', epoch, logs)

    def on_batch_begin(self, mode, step=None, logs=None):
        self._check_mode(mode)
        name = 'on_{}_batch_begin'.format(mode)
        self._call(name, step, logs)

    def on_batch_end(self, mode, step=None, logs=None):
        self._check_mode(mode)
        name = 'on_{}_batch_end'.format(mode)
        self._call(name, step, logs)


class Callback(object):
    """
    Base class used to build new callbacks.

    Examples:

        .. code-block:: python
            
138
            import paddle
139 140

            # build a simple model checkpoint callback
141
            class ModelCheckpoint(paddle.callbacks.Callback):
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164
                def __init__(self, save_freq=1, save_dir=None):
                    self.save_freq = save_freq
                    self.save_dir = save_dir

                def on_epoch_end(self, epoch, logs=None):
                    if self.model is not None and epoch % self.save_freq == 0:
                        path = '{}/{}'.format(self.save_dir, epoch)
                        print('save checkpoint at {}'.format(path))
                        self.model.save(path)

    """

    def __init__(self):
        self.model = None
        self.params = {}

    def set_params(self, params):
        """
        Set parameters, which is dict. The keys contain:

          - 'batch_size': an integer. Number of samples per batch.
          - 'epochs': an integer. Number of epochs.
          - 'steps': an integer. Number of steps of one epoch.
165 166
          - 'verbose': an integer. Verbose mode is 0, 1 or 2. 0 = silent, 1 = progress bar, 2 = one line per epoch.
          - 'metrics': a list of str. Names of metrics, including 'loss' and the names of paddle.metric.Metric.
167 168 169 170
        """
        self.params = params

    def set_model(self, model):
171
        """model is instance of paddle.Model.
172 173 174 175 176 177 178 179 180 181 182 183 184 185 186
        """
        self.model = model

    def on_train_begin(self, logs=None):
        """Called at the start of training.

        Args:
            logs (dict): The logs is a dict or None.
        """

    def on_train_end(self, logs=None):
        """Called at the end of training.

        Args:
            logs (dict): The logs is a dict or None. The keys of logs
187
                passed by paddle.Model contains 'loss', metric names and
188 189 190 191 192 193 194 195
                `batch_size`.
        """

    def on_eval_begin(self, logs=None):
        """Called at the start of evaluation.

        Args:
            logs (dict): The logs is a dict or None. The keys of logs
196
                passed by paddle.Model contains 'steps' and 'metrics',
197 198
                The `steps` is number of total steps of validation dataset.
                The `metrics` is a list of str including 'loss' and the names
199
                of paddle.metric.Metric.
200 201 202 203 204 205 206
        """

    def on_eval_end(self, logs=None):
        """Called at the end of evaluation.

        Args:
            logs (dict): The logs is a dict or None. The `logs` passed by
207
                paddle.Model is a dict contains 'loss', metrics and 'batch_size'
208 209 210
                of last batch of validation dataset.
        """

211
    def on_predict_begin(self, logs=None):
212 213 214 215 216 217
        """Called at the beginning of predict.

        Args:
            logs (dict): The logs is a dict or None.
        """

218
    def on_predict_end(self, logs=None):
219 220 221 222 223 224 225 226 227 228 229 230
        """Called at the end of predict.

        Args:
            logs (dict): The logs is a dict or None.
        """

    def on_epoch_begin(self, epoch, logs=None):
        """Called at the beginning of each epoch.

        Args:
            epoch (int): The index of epoch.
            logs (dict): The logs is a dict or None. The `logs` passed by
231
                paddle.Model is None.
232 233 234 235 236 237 238 239
        """

    def on_epoch_end(self, epoch, logs=None):
        """Called at the end of each epoch.

        Args:
            epoch (int): The index of epoch.
            logs (dict): The logs is a dict or None. The `logs` passed by
240
                paddle.Model is a dict, contains 'loss', metrics and 'batch_size'
241 242 243 244 245 246 247 248 249
                of last batch.
        """

    def on_train_batch_begin(self, step, logs=None):
        """Called at the beginning of each batch in training.

        Args:
            step (int): The index of step (or iteration).
            logs (dict): The logs is a dict or None. The `logs` passed by
250
                paddle.Model is empty.
251 252 253 254 255 256 257 258
        """

    def on_train_batch_end(self, step, logs=None):
        """Called at the end of each batch in training.

        Args:
            step (int): The index of step (or iteration).
            logs (dict): The logs is a dict or None. The `logs` passed by
259
                paddle.Model is a dict, contains 'loss', metrics and 'batch_size'
260 261 262 263 264 265 266 267 268
                of current batch.
        """

    def on_eval_batch_begin(self, step, logs=None):
        """Called at the beginning of each batch in evaluation.

        Args:
            step (int): The index of step (or iteration).
            logs (dict): The logs is a dict or None. The `logs` passed by
269
                paddle.Model is empty.
270 271 272 273 274 275 276 277
        """

    def on_eval_batch_end(self, step, logs=None):
        """Called at the end of each batch in evaluation.

        Args:
            step (int): The index of step (or iteration).
            logs (dict): The logs is a dict or None. The `logs` passed by
278
                paddle.Model is a dict, contains 'loss', metrics and 'batch_size'
279 280 281
                of current batch.
        """

282
    def on_predict_batch_begin(self, step, logs=None):
283 284 285 286 287 288 289
        """Called at the beginning of each batch in predict.

        Args:
            step (int): The index of step (or iteration).
            logs (dict): The logs is a dict or None.
        """

290
    def on_predict_batch_end(self, step, logs=None):
291 292 293 294 295 296 297 298 299
        """Called at the end of each batch in predict.

        Args:
            step (int): The index of step (or iteration).
            logs (dict): The logs is a dict or None.
        """


class ProgBarLogger(Callback):
300 301 302
    """
    Logger callback function.

303
    Args:
304 305
        log_freq (int): The frequency, in number of steps,
            the logs such as loss, metrics are printed. Default: 1.
306
        verbose (int): The verbosity mode, should be 0, 1, or 2.
307 308 309
            0 = silent, 1 = progress bar, 2 = one line per epoch, 3 = 2 + 
            time counter, such as average reader cost, samples per second. 
            Default: 2.
310 311 312 313

    Examples:
        .. code-block:: python

314
            import paddle
315
            import paddle.vision.transforms as T
316
            from paddle.vision.datasets import MNIST
317
            from paddle.static import InputSpec
318

319 320
            inputs = [InputSpec([-1, 1, 28, 28], 'float32', 'image')]
            labels = [InputSpec([None, 1], 'int64', 'label')]
321

322 323 324 325
            transform = T.Compose([
                T.Transpose(),
                T.Normalize([127.5], [127.5])
            ])
326
            train_dataset = MNIST(mode='train', transform=transform)
327

L
LielinJiang 已提交
328 329
            lenet = paddle.vision.LeNet()
            model = paddle.Model(lenet,
330
                inputs, labels)
331

L
LielinJiang 已提交
332
            optim = paddle.optimizer.Adam(0.001, parameters=lenet.parameters())
333
            model.prepare(optimizer=optim,
334 335
                        loss=paddle.nn.CrossEntropyLoss(),
                        metrics=paddle.metric.Accuracy())
336

337
            callback = paddle.callbacks.ProgBarLogger(log_freq=10)
338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356
            model.fit(train_dataset, batch_size=64, callbacks=callback)
    """

    def __init__(self, log_freq=1, verbose=2):
        self.epochs = None
        self.steps = None
        self.progbar = None
        self.verbose = verbose
        self.log_freq = log_freq

    def _is_print(self):
        return self.verbose and ParallelEnv().local_rank == 0

    def on_train_begin(self, logs=None):
        self.epochs = self.params['epochs']
        assert self.epochs
        self.train_metrics = self.params['metrics']
        assert self.train_metrics

357 358 359 360 361 362 363 364 365 366 367
        self._train_timer = {
            'data_time': 0,
            'batch_time': 0,
            'count': 0,
            'samples': 0,
        }
        if self._is_print():
            print(
                "The loss value printed in the log is the current batch, and the metric is the average value of previous step."
            )

368 369 370 371 372 373 374 375
    def on_epoch_begin(self, epoch=None, logs=None):
        self.steps = self.params['steps']
        self.epoch = epoch
        self.train_step = 0
        if self.epochs and self._is_print():
            print('Epoch %d/%d' % (epoch + 1, self.epochs))
        self.train_progbar = ProgressBar(num=self.steps, verbose=self.verbose)

376 377
        self._train_timer['batch_start_time'] = time.time()

378 379 380 381 382 383 384 385 386 387
    def _updates(self, logs, mode):
        values = []
        metrics = getattr(self, '%s_metrics' % (mode))
        progbar = getattr(self, '%s_progbar' % (mode))
        steps = getattr(self, '%s_step' % (mode))

        for k in metrics:
            if k in logs:
                values.append((k, logs[k]))

388 389 390 391 392 393 394 395 396 397 398 399
        if self.verbose == 3 and hasattr(self, '_%s_timer' % (mode)):
            timer = getattr(self, '_%s_timer' % (mode))
            cnt = timer['count'] if timer['count'] > 0 else 1.0
            samples = timer['samples'] if timer['samples'] > 0 else 1.0
            values.append(
                ('avg_reader_cost', "%.5f sec" % (timer['data_time'] / cnt)))
            values.append(
                ('avg_batch_cost', "%.5f sec" % (timer['batch_time'] / cnt)))
            values.append(
                ('ips', "%.5f samples/sec" %
                 (samples / (timer['batch_time'] + timer['batch_time']))))

400 401
        progbar.update(steps, values)

402 403 404 405 406 407
    def on_train_batch_begin(self, step, logs=None):
        self._train_timer['batch_data_end_time'] = time.time()
        self._train_timer['data_time'] += (
            self._train_timer['batch_data_end_time'] -
            self._train_timer['batch_start_time'])

408 409 410 411
    def on_train_batch_end(self, step, logs=None):
        logs = logs or {}
        self.train_step += 1

412 413 414 415 416
        self._train_timer['batch_time'] += (
            time.time() - self._train_timer['batch_data_end_time'])
        self._train_timer['count'] += 1
        samples = logs.get('batch_size', 1)
        self._train_timer['samples'] += samples
417 418 419
        if self._is_print() and self.train_step % self.log_freq == 0:
            if self.steps is None or self.train_step < self.steps:
                self._updates(logs, 'train')
420
        self._train_timer['batch_start_time'] = time.time()
421 422 423 424 425 426 427 428 429 430 431 432

    def on_epoch_end(self, epoch, logs=None):
        logs = logs or {}
        if self._is_print() and (self.steps is not None):
            self._updates(logs, 'train')

    def on_eval_begin(self, logs=None):
        self.eval_steps = logs.get('steps', None)
        self.eval_metrics = logs.get('metrics', [])
        self.eval_step = 0
        self.evaled_samples = 0

433 434 435 436 437 438 439
        self._eval_timer = {
            'data_time': 0,
            'batch_time': 0,
            'count': 0,
            'samples': 0,
        }

440 441 442 443
        self.eval_progbar = ProgressBar(
            num=self.eval_steps, verbose=self.verbose)
        if self._is_print():
            print('Eval begin...')
444 445 446 447 448 449 450 451 452 453 454
            print(
                "The loss value printed in the log is the current batch, and the metric is the average value of previous step."
            )

        self._eval_timer['batch_start_time'] = time.time()

    def on_eval_batch_begin(self, step, logs=None):
        self._eval_timer['batch_data_end_time'] = time.time()
        self._eval_timer['data_time'] += (
            self._eval_timer['batch_data_end_time'] -
            self._eval_timer['batch_start_time'])
455 456 457 458 459 460 461

    def on_eval_batch_end(self, step, logs=None):
        logs = logs or {}
        self.eval_step += 1
        samples = logs.get('batch_size', 1)
        self.evaled_samples += samples

462 463 464 465 466 467
        self._eval_timer['batch_time'] += (
            time.time() - self._eval_timer['batch_data_end_time'])
        self._eval_timer['count'] += 1
        samples = logs.get('batch_size', 1)
        self._eval_timer['samples'] += samples

468 469 470 471
        if self._is_print() and self.eval_step % self.log_freq == 0:
            if self.eval_steps is None or self.eval_step < self.eval_steps:
                self._updates(logs, 'eval')

472 473 474
        self._eval_timer['batch_start_time'] = time.time()

    def on_predict_begin(self, logs=None):
475 476 477 478
        self.test_steps = logs.get('steps', None)
        self.test_metrics = logs.get('metrics', [])
        self.test_step = 0
        self.tested_samples = 0
479 480 481 482 483 484 485 486

        self._test_timer = {
            'data_time': 0,
            'batch_time': 0,
            'count': 0,
            'samples': 0,
        }

487 488 489 490 491
        self.test_progbar = ProgressBar(
            num=self.test_steps, verbose=self.verbose)
        if self._is_print():
            print('Predict begin...')

492 493 494 495 496 497 498 499 500
        self._test_timer['batch_start_time'] = time.time()

    def on_predict_batch_begin(self, step, logs=None):
        self._test_timer['batch_data_end_time'] = time.time()
        self._test_timer['data_time'] += (
            self._test_timer['batch_data_end_time'] -
            self._test_timer['batch_start_time'])

    def on_predict_batch_end(self, step, logs=None):
501 502 503 504 505
        logs = logs or {}
        self.test_step += 1
        samples = logs.get('batch_size', 1)
        self.tested_samples += samples

506 507 508 509 510 511
        self._test_timer['batch_time'] += (
            time.time() - self._test_timer['batch_data_end_time'])
        self._test_timer['count'] += 1
        samples = logs.get('batch_size', 1)
        self._test_timer['samples'] += samples

512 513 514 515
        if self.test_step % self.log_freq == 0 and self._is_print():
            if self.test_steps is None or self.test_step < self.test_steps:
                self._updates(logs, 'test')

516 517
        self._test_timer['batch_start_time'] = time.time()

518 519 520 521 522 523
    def on_eval_end(self, logs=None):
        logs = logs or {}
        if self._is_print() and (self.eval_steps is not None):
            self._updates(logs, 'eval')
            print('Eval samples: %d' % (self.evaled_samples))

524
    def on_predict_end(self, logs=None):
525 526 527 528 529 530 531 532
        logs = logs or {}
        if self._is_print():
            if self.test_step % self.log_freq != 0 or self.verbose == 1:
                self._updates(logs, 'test')
            print('Predict samples: %d' % (self.tested_samples))


class ModelCheckpoint(Callback):
533 534 535
    """
    Model checkpoint callback function.

536
    Args:
537 538
        save_freq(int): The frequency, in number of epochs, the model checkpoint
            are saved. Default: 1.
539
        save_dir(str|None): The directory to save checkpoint during training.
540
            If None, will not save checkpoint. Default: None.
541 542 543 544

    Examples:
        .. code-block:: python

545
            import paddle
546
            import paddle.vision.transforms as T
547
            from paddle.vision.datasets import MNIST
548
            from paddle.static import InputSpec
549

550 551
            inputs = [InputSpec([-1, 1, 28, 28], 'float32', 'image')]
            labels = [InputSpec([None, 1], 'int64', 'label')]
552

553 554 555 556
            transform = T.Compose([
                T.Transpose(),
                T.Normalize([127.5], [127.5])
            ])
557
            train_dataset = MNIST(mode='train', transform=transform)
558

L
LielinJiang 已提交
559 560
            lenet = paddle.vision.LeNet()
            model = paddle.Model(lenet,
561
                inputs, labels)
562

L
LielinJiang 已提交
563
            optim = paddle.optimizer.Adam(0.001, parameters=lenet.parameters())
564
            model.prepare(optimizer=optim,
565 566
                        loss=paddle.nn.CrossEntropyLoss(),
                        metrics=paddle.metric.Accuracy())
567

568
            callback = paddle.callbacks.ModelCheckpoint(save_dir='./temp')
569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584
            model.fit(train_dataset, batch_size=64, callbacks=callback)
    """

    def __init__(self, save_freq=1, save_dir=None):
        self.save_freq = save_freq
        self.save_dir = save_dir

    def on_epoch_begin(self, epoch=None, logs=None):
        self.epoch = epoch

    def _is_save(self):
        return self.model and self.save_dir and ParallelEnv().local_rank == 0

    def on_epoch_end(self, epoch, logs=None):
        if self._is_save() and self.epoch % self.save_freq == 0:
            path = '{}/{}'.format(self.save_dir, epoch)
585
            print('save checkpoint at {}'.format(os.path.abspath(path)))
586 587 588 589 590
            self.model.save(path)

    def on_train_end(self, logs=None):
        if self._is_save():
            path = '{}/final'.format(self.save_dir)
591
            print('save checkpoint at {}'.format(os.path.abspath(path)))
592
            self.model.save(path)
593 594


595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684
class LRScheduler(Callback):
    """Lr scheduler callback function
    Args:
        by_step(bool, optional): whether to update learning rate scheduler 
            by step. Default: True.
        by_epoch(bool, optional): whether to update learning rate scheduler 
            by epoch. Default: False.

    Examples:
        .. code-block:: python

            import paddle
            import paddle.vision.transforms as T
            from paddle.static import InputSpec

            inputs = [InputSpec([-1, 1, 28, 28], 'float32', 'image')]
            labels = [InputSpec([None, 1], 'int64', 'label')]

            transform = T.Compose([
                T.Transpose(),
                T.Normalize([127.5], [127.5])
            ])
            train_dataset = paddle.vision.datasets.MNIST(mode='train', transform=transform)

            lenet = paddle.vision.LeNet()
            model = paddle.Model(lenet,
                inputs, labels)

            base_lr = 1e-3
            boundaries = [5, 8]
            wamup_steps = 4
            
            def make_optimizer(parameters=None):
                momentum = 0.9
                weight_decay = 5e-4
                values = [base_lr * (0.1**i) for i in range(len(boundaries) + 1)]
                learning_rate = paddle.optimizer.lr.PiecewiseDecay(
                    boundaries=boundaries, values=values)
                learning_rate = paddle.optimizer.lr.LinearWarmup(
                    learning_rate=learning_rate,
                    warmup_steps=wamup_epochs,
                    start_lr=base_lr / 5.,
                    end_lr=base_lr,
                    verbose=True)
                optimizer = paddle.optimizer.Momentum(
                    learning_rate=learning_rate,
                    weight_decay=weight_decay,
                    momentum=momentum,
                    parameters=parameters)
                return optimizer
                
            optim = make_optimizer(parameters=lenet.parameters())
            model.prepare(optimizer=optim,
                        loss=paddle.nn.CrossEntropyLoss(),
                        metrics=paddle.metric.Accuracy())

            # if LRScheduler callback not set, an instance LRScheduler update by step 
            # will be created auto.
            model.fit(train_dataset, batch_size=64)

            # create a learning rate scheduler update by epoch
            callback = paddle.callbacks.LRScheduler(by_step=False, by_epoch=True)
            model.fit(train_dataset, batch_size=64, callbacks=callback)
    """

    def __init__(self, by_step=True, by_epoch=False):
        if by_step and by_epoch:
            raise ValueError(
                "by_step option is mutually exclusive with by_epoch")

        self.by_step = by_step
        self.by_epoch = by_epoch

    def on_epoch_end(self, epoch, logs=None):
        if self.by_epoch:
            if self.model._optimizer and \
                hasattr(self.model._optimizer, '_learning_rate') and \
                isinstance(self.model._optimizer._learning_rate,
                           paddle.optimizer.lr.LRScheduler):
                self.model._optimizer._learning_rate.step()

    def on_train_batch_end(self, step, logs=None):
        if self.by_step:
            if self.model._optimizer and \
                hasattr(self.model._optimizer, '_learning_rate') and \
                isinstance(self.model._optimizer._learning_rate,
                           paddle.optimizer.lr.LRScheduler):
                self.model._optimizer._learning_rate.step()


L
LiuChiachi 已提交
685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835
class EarlyStopping(Callback):
    """Stop training when the given monitor stopped improving during evaluation.
    Args:
        monitor(str): Quantity to be monitored. Default: 'loss'.
        mode(str|None): Mode should be one of 'auto', 'min' or 'max'. In 'min'
            mode, training will stop until monitored quantity stops decreasing.
            In 'max' mode, training will stop until monitored quantity stops
            increasing. In 'auto' mode, exact mode can be inferred by the name
            of monitor. If 'acc' in monitor, the mode will be considered as
            'max', otherwise the mode will be set to 'min'. Default: 'auto'.
        patience(int): Number of epochs with no improvement after which
            training will be stopped. Default: 0.
        verbose(int): The verbosity mode, should be 0 or 1. When verbose=0,
            logs will not be printed. When verbose=1, logs will be printed.
            Default: 1.
        min_delta(int|float): The minimum change of monitored quantity. If
            the change is less than min_delta, model could be considered as no
            improvement. Default: 0.
        baseline(int|float|None): Baseline value for the monitored quantity.
            Training will stop if the model doesn't show improvement over the
            baseline. Default: None.
        save_best_model(bool): Whether to save best model. Default: True.
        
    Examples:
        .. code-block:: python

            import paddle
            from paddle import Model
            from paddle.static import InputSpec
            from paddle.vision.models import LeNet
            from paddle.vision.datasets import MNIST
            from paddle.metric import Accuracy
            from paddle.nn.layer.loss import CrossEntropyLoss
            import paddle.vision.transforms as T

            device = paddle.set_device('cpu')
            sample_num = 200
            save_dir = './best_model_checkpoint'
            transform = T.Compose(
                [T.Transpose(), T.Normalize([127.5], [127.5])])
            train_dataset = MNIST(mode='train', transform=transform)
            val_dataset = MNIST(mode='test', transform=transform)
            net = LeNet()
            optim = paddle.optimizer.Adam(
                learning_rate=0.001, parameters=net.parameters())

            inputs = [InputSpec([None, 1, 28, 28], 'float32', 'x')]
            labels = [InputSpec([None, 1], 'int64', 'label')]

            model = Model(net, inputs=inputs, labels=labels)
            model.prepare(
                optim,
                loss=CrossEntropyLoss(reduction="sum"),
                metrics=[Accuracy()])
            callbacks = paddle.callbacks.EarlyStopping(
                'loss',
                mode='min',
                patience=1,
                verbose=1,
                min_delta=0,
                baseline=None,
                save_best_model=True)
            model.fit(train_dataset,
                      val_dataset,
                      batch_size=64,
                      log_freq=200,
                      save_freq=10,
                      save_dir=save_dir,
                      epochs=20,
                      callbacks=[callbacks])
    """

    def __init__(self,
                 monitor='loss',
                 mode='auto',
                 patience=0,
                 verbose=1,
                 min_delta=0,
                 baseline=None,
                 save_best_model=True):
        super(EarlyStopping, self).__init__()
        self.monitor = monitor
        self.patience = patience
        self.verbose = verbose
        self.baseline = baseline
        self.min_delta = abs(min_delta)
        self.wait_epoch = 0
        self.best_weights = None
        self.stopped_epoch = 0
        self.save_best_model = save_best_model
        self.save_dir = None  # `save_dir` is get from `config_callbacks`
        if mode not in ['auto', 'min', 'max']:
            warnings.warn('EarlyStopping mode %s is unknown, '
                          'fallback to auto mode.' % mode)
            mode = 'auto'
        if mode == 'min':
            self.monitor_op = np.less
        elif mode == 'max':
            self.monitor_op = np.greater
        # When mode == 'auto', the mode should be inferred by `self.monitor`
        else:
            if 'acc' in self.monitor:
                self.monitor_op = np.greater
            else:
                self.monitor_op = np.less

        if self.monitor_op == np.greater:
            self.min_delta *= 1
        else:
            self.min_delta *= -1

    def on_train_begin(self, logs=None):
        self.wait_epoch = 0
        if self.baseline is not None:
            self.best_value = self.baseline
        else:
            self.best_value = np.inf if self.monitor_op == np.less else -np.inf
            self.best_weights = None

    def on_eval_end(self, logs=None):
        if logs is None or self.monitor not in logs:
            warnings.warn(
                'Monitor of EarlyStopping should be loss or metric name.')
            return
        current = logs[self.monitor]
        if isinstance(current, (list, tuple)):
            current = current[0]
        elif isinstance(current, numbers.Number):
            current = current
        else:
            return

        if self.monitor_op(current - self.min_delta, self.best_value):
            self.best_value = current
            self.wait_epoch = 0
            if self.save_best_model and self.save_dir is not None:
                path = os.path.join(self.save_dir, 'best_model')
                self.model.save(path)
        else:
            self.wait_epoch += 1
        if self.wait_epoch >= self.patience:
            self.model.stop_training = True
            if self.verbose > 0:
                print('Epoch %d: Early stopping.' % (self.stopped_epoch + 1))
                if self.save_best_model and self.save_dir is not None:
                    print('Best checkpoint has been saved at %s' %
                          (os.path.abspath(
                              os.path.join(self.save_dir, 'best_model'))))
        self.stopped_epoch += 1


836
class VisualDL(Callback):
837 838 839
    """
    VisualDL callback function.

840 841 842 843 844 845 846
    Args:
        log_dir (str): The directory to save visualdl log file.

    Examples:
        .. code-block:: python

            import paddle
847
            import paddle.vision.transforms as T
848 849 850 851 852
            from paddle.static import InputSpec

            inputs = [InputSpec([-1, 1, 28, 28], 'float32', 'image')]
            labels = [InputSpec([None, 1], 'int64', 'label')]

853 854 855 856 857 858
            transform = T.Compose([
                T.Transpose(),
                T.Normalize([127.5], [127.5])
            ])
            train_dataset = paddle.vision.datasets.MNIST(mode='train', transform=transform)
            eval_dataset = paddle.vision.datasets.MNIST(mode='test', transform=transform)
859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948

            net = paddle.vision.LeNet()
            model = paddle.Model(net, inputs, labels)

            optim = paddle.optimizer.Adam(0.001, parameters=net.parameters())
            model.prepare(optimizer=optim,
                        loss=paddle.nn.CrossEntropyLoss(),
                        metrics=paddle.metric.Accuracy())
            
            ## uncomment following lines to fit model with visualdl callback function
            # callback = paddle.callbacks.VisualDL(log_dir='visualdl_log_dir')
            # model.fit(train_dataset, eval_dataset, batch_size=64, callbacks=callback)

    """

    def __init__(self, log_dir):
        self.log_dir = log_dir
        self.epochs = None
        self.steps = None
        self.epoch = 0

    def _is_write(self):
        return ParallelEnv().local_rank == 0

    def on_train_begin(self, logs=None):
        self.epochs = self.params['epochs']
        assert self.epochs
        self.train_metrics = self.params['metrics']
        assert self.train_metrics
        self._is_fit = True
        self.train_step = 0

    def on_epoch_begin(self, epoch=None, logs=None):
        self.steps = self.params['steps']
        self.epoch = epoch

    def _updates(self, logs, mode):
        if not self._is_write():
            return
        if not hasattr(self, 'writer'):
            visualdl = try_import('visualdl')
            self.writer = visualdl.LogWriter(self.log_dir)

        metrics = getattr(self, '%s_metrics' % (mode))
        current_step = getattr(self, '%s_step' % (mode))

        if mode == 'train':
            total_step = current_step
        else:
            total_step = self.epoch

        for k in metrics:
            if k in logs:
                temp_tag = mode + '/' + k

                if isinstance(logs[k], (list, tuple)):
                    temp_value = logs[k][0]
                elif isinstance(logs[k], numbers.Number):
                    temp_value = logs[k]
                else:
                    continue

                self.writer.add_scalar(
                    tag=temp_tag, step=total_step, value=temp_value)

    def on_train_batch_end(self, step, logs=None):
        logs = logs or {}
        self.train_step += 1

        if self._is_write():
            self._updates(logs, 'train')

    def on_eval_begin(self, logs=None):
        self.eval_steps = logs.get('steps', None)
        self.eval_metrics = logs.get('metrics', [])
        self.eval_step = 0
        self.evaled_samples = 0

    def on_train_end(self, logs=None):
        if hasattr(self, 'writer'):
            self.writer.close()
            delattr(self, 'writer')

    def on_eval_end(self, logs=None):
        if self._is_write():
            self._updates(logs, 'eval')

            if (not hasattr(self, '_is_fit')) and hasattr(self, 'writer'):
                self.writer.close()
                delattr(self, 'writer')
L
LielinJiang 已提交
949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116


class ReduceLROnPlateau(Callback):
    """Reduce learning rate when a metric of evaluation has stopped improving.
    Models often benefit from reducing the learning rate by a factor
    of 2-10 once learning stagnates. This callback monitors a
    quantity and if no improvement is seen for a 'patience' number
    of epochs, the learning rate is reduced.
    
    Args:
        monitor(str, optional): Quantity to be monitored. Default: 'loss'.
        factor(float, optional): factor by which the learning rate will be reduced.
            `new_lr = lr * factor`. Default: 0.1.
        patience(int, optional): Number of epochs with no improvement after which
            learning rate will be reduced. Default: 10.
        verbose(int, optional): The verbosity mode. 0: quiet, 1: update messages.
            Default: 1.
        mode(str, optional): one of `{'auto', 'min', 'max'}`. In `'min'` mode,
            the learning rate will be reduced when the quantity monitored has 
            stopped decreasing. In 'max' mode, learning rate will reduce until 
            monitored quantity stops increasing. In 'auto' mode, exact mode 
            can be inferred by the name of monitor. If 'acc' in monitor, the 
            mode will be considered as 'max', otherwise the mode will be set 
            to 'min'. Default: 'auto'.
        min_delta(int|float, optional): threshold for measuring the new optimum, 
            to only focus on significant changes. Default: 0.
        cooldown(int, optional): number of epochs to wait before resuming normal operation after
            lr has been reduced. Default: 0.
        min_lr(float, optional): lower bound on the learning rate. Default: 0.
  
    Examples:
          .. code-block:: python
  
              import paddle
              from paddle import Model
              from paddle.static import InputSpec
              from paddle.vision.models import LeNet
              from paddle.vision.datasets import MNIST
              from paddle.metric import Accuracy
              from paddle.nn.layer.loss import CrossEntropyLoss
              import paddle.vision.transforms as T  
              sample_num = 200
              transform = T.Compose(
                  [T.Transpose(), T.Normalize([127.5], [127.5])])
              train_dataset = MNIST(mode='train', transform=transform)
              val_dataset = MNIST(mode='test', transform=transform)
              net = LeNet()
              optim = paddle.optimizer.Adam(
                  learning_rate=0.001, parameters=net.parameters())  
              inputs = [InputSpec([None, 1, 28, 28], 'float32', 'x')]
              labels = [InputSpec([None, 1], 'int64', 'label')]  
              model = Model(net, inputs=inputs, labels=labels)
              model.prepare(
                  optim,
                  loss=CrossEntropyLoss(),
                  metrics=[Accuracy()])  
              callbacks = paddle.callbacks.ReduceLROnPlateau(patience=3, verbose=1)
              model.fit(train_dataset,
                          val_dataset,
                          batch_size=64,
                          log_freq=200,
                          save_freq=10,
                          epochs=20,
                          callbacks=[callbacks])
  
    """

    def __init__(self,
                 monitor='loss',
                 factor=0.1,
                 patience=10,
                 verbose=1,
                 mode='auto',
                 min_delta=1e-4,
                 cooldown=0,
                 min_lr=0):
        super(ReduceLROnPlateau, self).__init__()

        self.monitor = monitor
        if factor >= 1.0:
            raise ValueError('ReduceLROnPlateau '
                             'does not support a factor >= 1.0.')

        self.factor = factor
        self.min_lr = min_lr
        self.min_delta = min_delta
        self.patience = patience
        self.verbose = verbose
        self.cooldown = cooldown
        self.cooldown_counter = 0  # Cooldown counter.
        self.wait = 0
        self.best = 0
        self.mode = mode
        self.monitor_op = None
        self.epoch = 0
        self._reset()

    def _reset(self):
        """Resets wait counter and cooldown counter.
        """
        if self.mode not in ['auto', 'min', 'max']:
            warnings.warn('Learning rate reduction mode %s is unknown, '
                          'fallback to auto mode.' % self.mode)
            self.mode = 'auto'
        if (self.mode == 'min' or
            (self.mode == 'auto' and 'acc' not in self.monitor)):
            self.monitor_op = lambda a, b: np.less(a, b - self.min_delta)
            self.best = np.Inf
        else:
            self.monitor_op = lambda a, b: np.greater(a, b + self.min_delta)
            self.best = -np.Inf
        self.cooldown_counter = 0
        self.wait = 0

    def on_train_begin(self, logs=None):
        self._reset()

    def on_eval_end(self, logs=None):
        if logs is None or self.monitor not in logs:
            warnings.warn(
                'Monitor of ReduceLROnPlateau should be loss or metric name.')
            return
        else:
            try:
                lr = self.model._optimizer._learning_rate
                if not isinstance(lr, float):
                    warnings.warn(
                        'Expected learning_rate be float, bug got {}.'.format(
                            type(lr)))
                    return
            except Exception as e:
                warnings.warn(
                    'There are something wrong when get learning_rate from optimizer: {}.'.
                    format(e))
                return

        current = logs[self.monitor]
        if isinstance(current, (list, tuple)):
            current = current[0]
        elif isinstance(current, numbers.Number):
            current = current
        else:
            return

        if self.in_cooldown():
            self.cooldown_counter -= 1
            self.wait = 0

        if self.monitor_op(current, self.best):
            self.best = current
            self.wait = 0
        elif not self.in_cooldown():
            self.wait += 1
            if self.wait >= self.patience:
                old_lr = self.model._optimizer.get_lr()
                if old_lr > np.float32(self.min_lr):
                    new_lr = old_lr * self.factor
                    new_lr = max(new_lr, self.min_lr)
                    self.model._optimizer._learning_rate = new_lr
                    if self.verbose > 0 and ParallelEnv().local_rank == 0:
                        print('\nEpoch %d: ReduceLROnPlateau reducing learning '
                              'rate to %s.' % (self.epoch + 1, new_lr))
                    self.cooldown_counter = self.cooldown
                    self.wait = 0
        self.epoch += 1

    def in_cooldown(self):
        return self.cooldown_counter > 0