model_training_utils.py 21 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# Copyright 2019 The TensorFlow 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
"""A light weight utilities to train NLP models."""
16
17
18
19
20

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

21
import json
22
import os
23
import tempfile
24
25

from absl import logging
Hongkun Yu's avatar
Hongkun Yu committed
26
import tensorflow as tf
Zongwei Zhou's avatar
Zongwei Zhou committed
27
from official.staging.training import grad_utils
28
from official.utils.misc import distribution_utils
29

30
31
_SUMMARY_TXT = 'training_summary.txt'
_MIN_SUMMARY_STEPS = 10
32

33

34
35
36
37
38
39
40
41
42
def _should_export_checkpoint(strategy):
  return (not strategy) or strategy.extended.should_checkpoint


def _should_export_summary(strategy):
  return (not strategy) or strategy.extended.should_save_summary


def _save_checkpoint(strategy, checkpoint, model_dir, checkpoint_prefix):
43
44
  """Saves model to with provided checkpoint prefix."""

45
46
47
48
49
50
51
52
53
54
55
56
  if _should_export_checkpoint(strategy):
    checkpoint_path = os.path.join(model_dir, checkpoint_prefix)
    saved_path = checkpoint.save(checkpoint_path)
    logging.info('Saving model as TF checkpoint: %s', saved_path)
  else:
    # In multi worker training we need every worker to save checkpoint, because
    # variables can trigger synchronization on read and synchronization needs
    # all workers to participate. To avoid workers overriding each other we save
    # to a temporary directory on non-chief workers.
    tmp_dir = tempfile.mkdtemp()
    checkpoint.save(os.path.join(tmp_dir, 'ckpt'))
    tf.io.gfile.rmtree(tmp_dir)
57
58
59
  return


60
61
62
63
64
def _get_input_iterator(input_fn, strategy):
  """Returns distributed dataset iterator."""
  # When training with TPU pods, datasets needs to be cloned across
  # workers. Since Dataset instance cannot be cloned in eager mode, we instead
  # pass callable that returns a dataset.
Hongkun Yu's avatar
Hongkun Yu committed
65
66
67
68
  if not callable(input_fn):
    raise ValueError('`input_fn` should be a closure that returns a dataset.')
  iterator = iter(
      strategy.experimental_distribute_datasets_from_function(input_fn))
69
70
71
  return iterator


72
73
74
75
76
def _float_metric_value(metric):
  """Gets the value of a float-value keras metric."""
  return metric.result().numpy().astype(float)


77
def steps_to_run(current_step, steps_per_epoch, steps_per_loop):
78
  """Calculates steps to run on device."""
79
80
81
  if steps_per_loop <= 0:
    raise ValueError('steps_per_loop should be positive integer.')
  if steps_per_loop == 1:
82
83
84
85
86
87
88
89
    return steps_per_loop
  remainder_in_epoch = current_step % steps_per_epoch
  if remainder_in_epoch != 0:
    return min(steps_per_epoch - remainder_in_epoch, steps_per_loop)
  else:
    return steps_per_loop


90
def write_txt_summary(training_summary, summary_dir):
91
  """Writes a summary text file to record stats."""
Chen Chen's avatar
Chen Chen committed
92
93
  if not tf.io.gfile.exists(summary_dir):
    tf.io.gfile.mkdir(summary_dir)
94
  summary_path = os.path.join(summary_dir, _SUMMARY_TXT)
95
96
97
98
99
  with tf.io.gfile.GFile(summary_path, 'wb') as f:
    logging.info('Training Summary: \n%s', str(training_summary))
    f.write(json.dumps(training_summary, indent=4))


100
101
102
103
104
105
106
def run_customized_training_loop(
    # pylint: disable=invalid-name
    _sentinel=None,
    # pylint: enable=invalid-name
    strategy=None,
    model_fn=None,
    loss_fn=None,
107
    scale_loss=True,
108
109
110
    model_dir=None,
    train_input_fn=None,
    steps_per_epoch=None,
111
    steps_per_loop=1,
112
113
114
115
116
    epochs=1,
    eval_input_fn=None,
    eval_steps=None,
    metric_fn=None,
    init_checkpoint=None,
117
    custom_callbacks=None,
Chen Chen's avatar
Chen Chen committed
118
    run_eagerly=False,
Zongwei Zhou's avatar
Zongwei Zhou committed
119
120
121
    sub_model_export_name=None,
    explicit_allreduce=False,
    pre_allreduce_callbacks=None,
Chen Chen's avatar
Chen Chen committed
122
123
    post_allreduce_callbacks=None,
    train_summary_interval=0):
124
125
126
127
128
129
130
131
132
133
134
135
  """Run BERT pretrain model training using low-level API.

  Arguments:
      _sentinel: Used to prevent positional parameters. Internal, do not use.
      strategy: Distribution strategy on which to run low level training loop.
      model_fn: Function that returns a tuple (model, sub_model). Caller of this
        function should add optimizer to the `model` via calling
        `model.compile()` API or manually setting `model.optimizer` attribute.
        Second element of the returned tuple(sub_model) is an optional sub model
        to be used for initial checkpoint -- if provided.
      loss_fn: Function with signature func(labels, logits) and returns a loss
        tensor.
136
137
      scale_loss: Whether to divide the raw loss by number of replicas before
        gradients calculation.
138
139
140
      model_dir: Model directory used during training for restoring/saving model
        weights.
      train_input_fn: Function that returns a tf.data.Dataset used for training.
141
142
143
144
145
146
      steps_per_epoch: Number of steps to run per epoch. At the end of each
        epoch, model checkpoint will be saved and evaluation will be conducted
        if evaluation dataset is provided.
      steps_per_loop: Number of steps per graph-mode loop. In order to reduce
        communication in eager context, training logs are printed every
        steps_per_loop.
147
148
149
150
151
152
153
154
155
156
      epochs: Number of epochs to train.
      eval_input_fn: Function that returns evaluation dataset. If none,
        evaluation is skipped.
      eval_steps: Number of steps to run evaluation. Required if `eval_input_fn`
        is not none.
      metric_fn: A metrics function that returns a Keras Metric object to record
        evaluation result using evaluation dataset or with training dataset
        after every epoch.
      init_checkpoint: Optional checkpoint to load to `sub_model` returned by
        `model_fn`.
157
      custom_callbacks: A list of Keras Callbacks objects to run during
158
        training. More specifically, `on_batch_begin()`, `on_batch_end()`,
159
160
        `on_epoch_begin()`, `on_epoch_end()` methods are invoked during
        training.  Note that some metrics may be missing from `logs`.
161
162
      run_eagerly: Whether to run model training in pure eager execution. This
        should be disable for TPUStrategy.
Chen Chen's avatar
Chen Chen committed
163
164
165
166
167
      sub_model_export_name: If not None, will export `sub_model` returned by
        `model_fn` into checkpoint files. The name of intermediate checkpoint
        file is {sub_model_export_name}_step_{step}.ckpt and the last
        checkpint's name is {sub_model_export_name}.ckpt;
        if None, `sub_model` will not be exported as checkpoint.
Zongwei Zhou's avatar
Zongwei Zhou committed
168
169
170
171
172
173
174
175
176
177
      explicit_allreduce: Whether to explicitly perform gradient allreduce,
        instead of relying on implicit allreduce in optimizer.apply_gradients().
        default is False. For now, if training using FP16 mixed precision,
        explicit allreduce will aggregate gradients in FP16 format. For TPU and
        GPU training using FP32, explicit allreduce will aggregate gradients in
        FP32 format.
      pre_allreduce_callbacks: A list of callback functions that takes gradients
        and model variables pairs as input, manipulate them, and returns a new
        gradients and model variables paris. The callback functions will be
        invoked in the list order and before gradients are allreduced.
178
179
180
        With mixed precision training, the pre_allreduce_allbacks will be
        applied on scaled_gradients. Default is no callbacks.
        Only used when explicit_allreduce=True.
Zongwei Zhou's avatar
Zongwei Zhou committed
181
182
183
184
185
186
      post_allreduce_callbacks: A list of callback functions that takes
        gradients and model variables pairs as input, manipulate them, and
        returns a new gradients and model variables paris. The callback
        functions will be invoked in the list order and right before gradients
        are applied to variables for updates. Default is no callbacks. Only used
        when explicit_allreduce=True.
Chen Chen's avatar
Chen Chen committed
187
188
      train_summary_interval: Step interval for training summaries. If the value
        is a negative number, then training summaries are not enabled.
189
190
191
192
193
194
195
196

  Returns:
      Trained model.

  Raises:
      ValueError: (1) When model returned by `model_fn` does not have optimizer
        attribute or when required parameters are set to none. (2) eval args are
        not specified correctly. (3) metric_fn must be a callable if specified.
Chen Chen's avatar
Chen Chen committed
197
198
        (4) sub_model_checkpoint_name is specified, but `sub_model` returned
        by `model_fn` is None.
199
200
201
202
203
204
205
206
207
208
209
  """

  if _sentinel is not None:
    raise ValueError('only call `run_customized_training_loop()` '
                     'with named arguments.')

  required_arguments = [
      strategy, model_fn, loss_fn, model_dir, steps_per_epoch, train_input_fn
  ]
  if [arg for arg in required_arguments if arg is None]:
    raise ValueError('`strategy`, `model_fn`, `loss_fn`, `model_dir`, '
210
211
212
213
214
215
216
217
                     '`steps_per_loop` and `steps_per_epoch` are required '
                     'parameters.')
  if steps_per_loop > steps_per_epoch:
    logging.error(
        'steps_per_loop: %d is specified to be greater than '
        ' steps_per_epoch: %d, we will use steps_per_epoch as'
        ' steps_per_loop.', steps_per_loop, steps_per_epoch)
    steps_per_loop = steps_per_epoch
218
219
  assert tf.executing_eagerly()

220
221
222
  if run_eagerly:
    if isinstance(strategy, tf.distribute.experimental.TPUStrategy):
      raise ValueError(
A. Unique TensorFlower's avatar
A. Unique TensorFlower committed
223
          'TPUStrategy should not run eagerly as it heavily relies on graph'
224
225
          ' optimization for the distributed system.')

226
227
228
229
230
231
232
233
  if eval_input_fn and (eval_steps is None or metric_fn is None):
    raise ValueError(
        '`eval_step` and `metric_fn` are required when `eval_input_fn ` '
        'is not none.')
  if metric_fn and not callable(metric_fn):
    raise ValueError(
        'if `metric_fn` is specified, metric_fn must be a callable.')

234
235
  callback_list = tf.keras.callbacks.CallbackList(custom_callbacks)

236
  total_training_steps = steps_per_epoch * epochs
237
238
239
240
241
242
243
244
245
  train_iterator = _get_input_iterator(train_input_fn, strategy)

  with distribution_utils.get_strategy_scope(strategy):
    # To correctly place the model weights on accelerators,
    # model and optimizer should be created in scope.
    model, sub_model = model_fn()
    if not hasattr(model, 'optimizer'):
      raise ValueError('User should set optimizer attribute to model '
                       'inside `model_fn`.')
Chen Chen's avatar
Chen Chen committed
246
247
248
249
    if sub_model_export_name and sub_model is None:
      raise ValueError('sub_model_export_name is specified as %s, but '
                       'sub_model is None.' % sub_model_export_name)

250
251
252
253
254
255
256
    optimizer = model.optimizer

    if init_checkpoint:
      logging.info(
          'Checkpoint file %s found and restoring from '
          'initial checkpoint for core model.', init_checkpoint)
      checkpoint = tf.train.Checkpoint(model=sub_model)
Jing Li's avatar
Jing Li committed
257
      checkpoint.restore(init_checkpoint).assert_existing_objects_matched()
258
259
260
261
262
263
264
265
266
267
268
269
270
      logging.info('Loading from checkpoint file completed')

    train_loss_metric = tf.keras.metrics.Mean(
        'training_loss', dtype=tf.float32)
    eval_metrics = [metric_fn()] if metric_fn else []
    # If evaluation is required, make a copy of metric as it will be used by
    # both train and evaluation.
    train_metrics = [
        metric.__class__.from_config(metric.get_config())
        for metric in eval_metrics
    ]

    # Create summary writers
271
272
273
274
275
276
277
    if _should_export_summary(strategy):
      summary_dir = os.path.join(model_dir, 'summaries')
    else:
      # In multi worker training we need every worker to write summary, because
      # variables can trigger synchronization on read and synchronization needs
      # all workers to participate.
      summary_dir = tempfile.mkdtemp()
278
    eval_summary_writer = tf.summary.create_file_writer(
279
        os.path.join(summary_dir, 'eval'))
Chen Chen's avatar
Chen Chen committed
280
281
    last_summary_step = 0
    if steps_per_loop >= _MIN_SUMMARY_STEPS and train_summary_interval >= 0:
282
283
284
      # Only writes summary when the stats are collected sufficiently over
      # enough steps.
      train_summary_writer = tf.summary.create_file_writer(
285
          os.path.join(summary_dir, 'train'))
286
    else:
Chen Chen's avatar
Chen Chen committed
287
      train_summary_writer = tf.summary.create_noop_writer()
288
289
290
291
292
293
294
295
296
297
298

    # Collects training variables.
    training_vars = model.trainable_variables

    def _replicated_step(inputs):
      """Replicated training step."""

      inputs, labels = inputs
      with tf.GradientTape() as tape:
        model_outputs = model(inputs, training=True)
        loss = loss_fn(labels, model_outputs)
299
300
301
302
303
304
        # Raw loss is used for reporting in metrics/logs.
        raw_loss = loss
        if scale_loss:
          # Scales down the loss for gradients to be invariant from replicas.
          loss = loss / strategy.num_replicas_in_sync

Zongwei Zhou's avatar
Zongwei Zhou committed
305
306
307
308
309
      if explicit_allreduce:
        grad_utils.minimize_using_explicit_allreduce(tape, optimizer, loss,
                                                     training_vars,
                                                     pre_allreduce_callbacks,
                                                     post_allreduce_callbacks)
310
      else:
Zongwei Zhou's avatar
Zongwei Zhou committed
311
312
313
314
315
316
317
318
319
        if isinstance(optimizer,
                      tf.keras.mixed_precision.experimental.LossScaleOptimizer):
          with tape:
            scaled_loss = optimizer.get_scaled_loss(loss)
          scaled_grads = tape.gradient(scaled_loss, training_vars)
          grads = optimizer.get_unscaled_gradients(scaled_grads)
        else:
          grads = tape.gradient(loss, training_vars)
        optimizer.apply_gradients(zip(grads, training_vars))
320
      # For reporting, the metric takes the mean of losses.
321
      train_loss_metric.update_state(raw_loss)
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
      for metric in train_metrics:
        metric.update_state(labels, model_outputs)

    @tf.function
    def train_steps(iterator, steps):
      """Performs distributed training steps in a loop.

      Args:
        iterator: the distributed iterator of training datasets.
        steps: an tf.int32 integer tensor to specify number of steps to run
          inside host training loop.

      Raises:
        ValueError: Any of the arguments or tensor shapes are invalid.
      """
      if not isinstance(steps, tf.Tensor):
        raise ValueError('steps should be an Tensor. Python object may cause '
                         'retracing.')

      for _ in tf.range(steps):
Ken Franko's avatar
Ken Franko committed
342
        strategy.run(_replicated_step, args=(next(iterator),))
343

344
345
    def train_single_step(iterator):
      """Performs a distributed training step.
346

347
348
      Args:
        iterator: the distributed iterator of training datasets.
349

350
351
352
      Raises:
        ValueError: Any of the arguments or tensor shapes are invalid.
      """
Ken Franko's avatar
Ken Franko committed
353
      strategy.run(_replicated_step, args=(next(iterator),))
354

355
356
    def test_step(iterator):
      """Calculates evaluation metrics on distributed devices."""
357

358
359
      def _test_step_fn(inputs):
        """Replicated accuracy calculation."""
360

361
362
363
364
        inputs, labels = inputs
        model_outputs = model(inputs, training=False)
        for metric in eval_metrics:
          metric.update_state(labels, model_outputs)
365

Ken Franko's avatar
Ken Franko committed
366
      strategy.run(_test_step_fn, args=(next(iterator),))
367
368
369
370
371
372

    if not run_eagerly:
      train_single_step = tf.function(train_single_step)
      test_step = tf.function(test_step)

    def _run_evaluation(current_training_step, test_iterator):
373
374
375
376
377
378
379
380
381
      """Runs validation steps and aggregate metrics.

      Args:
        current_training_step: tf.int32 tensor containing the current step.
        test_iterator: distributed iterator of test datasets.

      Returns:
        A dict of metic names and values.
      """
382
383
384
      for _ in range(eval_steps):
        test_step(test_iterator)

385
      logs = {}
386
387
388
      with eval_summary_writer.as_default():
        for metric in eval_metrics + model.metrics:
          metric_value = _float_metric_value(metric)
389
          logs[metric.name] = metric_value
390
391
392
393
394
395
          logging.info('Step: [%d] Validation %s = %f', current_training_step,
                       metric.name, metric_value)
          tf.summary.scalar(
              metric.name, metric_value, step=current_training_step)
        eval_summary_writer.flush()

396
      return logs
397
398

    # Training loop starts here.
Le Hou's avatar
Le Hou committed
399
400
    checkpoint = tf.train.Checkpoint(
        model=model, optimizer=optimizer, global_step=optimizer.iterations)
Chen Chen's avatar
Chen Chen committed
401
    sub_model_checkpoint = tf.train.Checkpoint(
Le Hou's avatar
Le Hou committed
402
403
        model=sub_model,
        global_step=optimizer.iterations) if sub_model_export_name else None
Chen Chen's avatar
Chen Chen committed
404

405
406
407
408
409
410
411
412
413
414
415
416
    latest_checkpoint_file = tf.train.latest_checkpoint(model_dir)
    if latest_checkpoint_file:
      logging.info(
          'Checkpoint file %s found and restoring from '
          'checkpoint', latest_checkpoint_file)
      checkpoint.restore(latest_checkpoint_file)
      logging.info('Loading from checkpoint file completed')

    current_step = optimizer.iterations.numpy()
    checkpoint_name = 'ctl_step_{step}.ckpt'

    while current_step < total_training_steps:
417
418
419
      if current_step % steps_per_epoch == 0:
        callback_list.on_epoch_begin(int(current_step / steps_per_epoch) + 1)

420
421
422
423
424
425
      # Training loss/metric are taking average over steps inside micro
      # training loop. We reset the their values before each round.
      train_loss_metric.reset_states()
      for metric in train_metrics + model.metrics:
        metric.reset_states()

426
      callback_list.on_batch_begin(current_step)
427
      # Runs several steps in the host while loop.
428
      steps = steps_to_run(current_step, steps_per_epoch, steps_per_loop)
429

430
      if tf.config.list_physical_devices('GPU'):
431
432
        # TODO(zongweiz): merge with train_steps once tf.while_loop
        # GPU performance bugs are fixed.
433
434
        for _ in range(steps):
          train_single_step(train_iterator)
435
436
437
438
      else:
        # Converts steps to a Tensor to avoid tf.function retracing.
        train_steps(train_iterator,
                    tf.convert_to_tensor(steps, dtype=tf.int32))
439
      train_loss = _float_metric_value(train_loss_metric)
440
      current_step += steps
441
      callback_list.on_batch_end(current_step - 1, {'loss': train_loss})
442
443
444
445
446

      # Updates training logging.
      training_status = 'Train Step: %d/%d  / loss = %s' % (
          current_step, total_training_steps, train_loss)

Chen Chen's avatar
Chen Chen committed
447
448
449
450
451
452
453
454
455
456
457
458
459
460
      if current_step >= last_summary_step + train_summary_interval:
        summary_writer = train_summary_writer
        last_summary_step = current_step
      else:
        summary_writer = tf.summary.create_noop_writer()

      with summary_writer.as_default():
        tf.summary.scalar(
            train_loss_metric.name, train_loss, step=current_step)
        for metric in train_metrics + model.metrics:
          metric_value = _float_metric_value(metric)
          training_status += '  %s = %f' % (metric.name, metric_value)
          tf.summary.scalar(metric.name, metric_value, step=current_step)
        summary_writer.flush()
461
462
463
      logging.info(training_status)

      if current_step % steps_per_epoch == 0:
464
465
466
467
468
469
470
471
472
        # Save a submodel with the step in the file name after each epoch.
        if sub_model_export_name:
          _save_checkpoint(
              strategy, sub_model_checkpoint, model_dir,
              '%s_step_%d.ckpt' % (sub_model_export_name, current_step))

        # Save model checkpoints and run validation steps after each epoch
        # (with the exception of the final epoch which is handled after the
        # training loop).
473
        if current_step < total_training_steps:
474
          _save_checkpoint(strategy, checkpoint, model_dir,
475
                           checkpoint_name.format(step=current_step))
476
          logs = None
477
478
          if eval_input_fn:
            logging.info('Running evaluation after step: %s.', current_step)
479
480
            logs = _run_evaluation(current_step,
                                   _get_input_iterator(eval_input_fn, strategy))
481
482
483
            # Re-initialize evaluation metric.
            for metric in eval_metrics + model.metrics:
              metric.reset_states()
484

485
486
          callback_list.on_epoch_end(int(current_step / steps_per_epoch), logs)

Chen Chen's avatar
Chen Chen committed
487
    if sub_model_export_name:
488
      _save_checkpoint(strategy, sub_model_checkpoint, model_dir,
Chen Chen's avatar
Chen Chen committed
489
                       '%s.ckpt' % sub_model_export_name)
490

491
492
    _save_checkpoint(strategy, checkpoint, model_dir,
                     checkpoint_name.format(step=current_step))
493
    logs = None
494
495
    if eval_input_fn:
      logging.info('Running final evaluation after training is complete.')
496
497
498
499
      logs = _run_evaluation(current_step,
                             _get_input_iterator(eval_input_fn, strategy))

    callback_list.on_epoch_end(int(current_step / steps_per_epoch), logs)
500

501
502
503
504
    training_summary = {
        'total_training_steps': total_training_steps,
        'train_loss': _float_metric_value(train_loss_metric),
    }
505
506
    for metric in model.metrics:
      training_summary[metric.name] = _float_metric_value(metric)
507
508
509
510
511
    if eval_metrics:
      # TODO(hongkuny): Cleans up summary reporting in text.
      training_summary['last_train_metrics'] = _float_metric_value(
          train_metrics[0])
      training_summary['eval_metrics'] = _float_metric_value(eval_metrics[0])
512

513
    write_txt_summary(training_summary, summary_dir)
514

515
516
517
    if not _should_export_summary(strategy):
      tf.io.gfile.rmtree(summary_dir)

518
    return model