resnet_run_loop.py 33.1 KB
Newer Older
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
# Copyright 2017 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.
# ==============================================================================
"""Contains utility and supporting functions for ResNet.

  This module contains ResNet code which does not directly build layers. This
includes dataset management, hyperparameter and optimizer code, and argument
parsing. Code for defining the ResNet layers can be found in resnet_model.py.
"""

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

26
import functools
Taylor Robie's avatar
Taylor Robie committed
27
import math
Toby Boyd's avatar
Toby Boyd committed
28
import multiprocessing
29
30
import os

31
# pylint: disable=g-bad-import-order
32
from absl import flags
33
import tensorflow as tf
34
35

from official.resnet import resnet_model
36
from official.utils.flags import core as flags_core
37
from official.utils.export import export
38
39
from official.utils.logs import hooks_helper
from official.utils.logs import logger
40
from official.resnet import imagenet_preprocessing
41
from official.utils.misc import distribution_utils
42
from official.utils.misc import model_helpers
43
44
45
46
47


################################################################################
# Functions for input processing.
################################################################################
Toby Boyd's avatar
Toby Boyd committed
48
49
50
51
52
53
54
55
def process_record_dataset(dataset,
                           is_training,
                           batch_size,
                           shuffle_buffer,
                           parse_record_fn,
                           num_epochs=1,
                           dtype=tf.float32,
                           datasets_num_private_threads=None,
56
57
                           num_parallel_batches=1,
                           drop_remainder=False):
Karmel Allison's avatar
Karmel Allison committed
58
  """Given a Dataset with raw records, return an iterator over the records.
59
60
61
62
63
64
65
66
67
68
69

  Args:
    dataset: A Dataset representing raw records
    is_training: A boolean denoting whether the input is for training.
    batch_size: The number of samples per batch.
    shuffle_buffer: The buffer size to use when shuffling records. A larger
      value results in better randomness, but smaller values reduce startup
      time and use less memory.
    parse_record_fn: A function that takes a raw record and returns the
      corresponding (image, label) pair.
    num_epochs: The number of epochs to repeat the dataset.
70
    dtype: Data type to use for images/features.
Toby Boyd's avatar
Toby Boyd committed
71
72
73
    datasets_num_private_threads: Number of threads for a private
      threadpool created for all datasets computation.
    num_parallel_batches: Number of parallel batches for tf.data.
74
75
    drop_remainder: A boolean indicates whether to drop the remainder of the
      batches. If True, the batch dimension will be static.
76
77
78
79

  Returns:
    Dataset of (image, label) pairs ready for iteration.
  """
80
81
82
83
84
85
86
87
  # Defines a specific size thread pool for tf.data operations.
  if datasets_num_private_threads:
    options = tf.data.Options()
    options.experimental_threading.private_threadpool_size = (
        datasets_num_private_threads)
    dataset = dataset.with_options(options)
    tf.compat.v1.logging.info('datasets_num_private_threads: %s',
                              datasets_num_private_threads)
88

Haoyu Zhang's avatar
Haoyu Zhang committed
89
90
91
92
93
  # Disable intra-op parallelism to optimize for throughput instead of latency.
  options = tf.data.Options()
  options.experimental_threading.max_intra_op_parallelism = 1
  dataset = dataset.with_options(options)

94
95
  # Prefetches a batch at a time to smooth out the time taken to load input
  # files for shuffling and processing.
96
97
  dataset = dataset.prefetch(buffer_size=batch_size)
  if is_training:
98
    # Shuffles records before repeating to respect epoch boundaries.
99
100
    dataset = dataset.shuffle(buffer_size=shuffle_buffer)

101
  # Repeats the dataset for the number of epochs to train.
102
103
  dataset = dataset.repeat(num_epochs)

104
  # Parses the raw records into images and labels.
Haoyu Zhang's avatar
Haoyu Zhang committed
105
106
107
  dataset = dataset.map(
      lambda value: parse_record_fn(value, is_training, dtype),
      num_parallel_calls=tf.data.experimental.AUTOTUNE)
108
  dataset = dataset.batch(batch_size, drop_remainder=drop_remainder)
109
110
111
112

  # Operations between the final prefetch and the get_next call to the iterator
  # will happen synchronously during run time. We prefetch here again to
  # background all of the above processing work and keep it out of the
113
114
115
  # critical training path. Setting buffer_size to tf.contrib.data.AUTOTUNE
  # allows DistributionStrategies to adjust how many batches to fetch based
  # on how many devices are present.
116
  dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
117
118
119
120

  return dataset


Toby Boyd's avatar
Toby Boyd committed
121
122
123
def get_synth_input_fn(height, width, num_channels, num_classes,
                       dtype=tf.float32):
  """Returns an input function that returns a dataset with random data.
124

Toby Boyd's avatar
Toby Boyd committed
125
126
127
128
  This input_fn returns a data set that iterates over a set of random data and
  bypasses all preprocessing, e.g. jpeg decode and copy. The host to device
  copy is still included. This used to find the upper throughput bound when
  tunning the full input pipeline.
129
130
131
132
133
134
135

  Args:
    height: Integer height that will be used to create a fake image tensor.
    width: Integer width that will be used to create a fake image tensor.
    num_channels: Integer depth that will be used to create a fake image tensor.
    num_classes: Number of classes that should be represented in the fake labels
      tensor
Toby Boyd's avatar
Toby Boyd committed
136
    dtype: Data type for features/images.
137
138
139
140
141

  Returns:
    An input_fn that can be used in place of a real one to return a dataset
    that can be used for iteration.
  """
Toby Boyd's avatar
Toby Boyd committed
142
143
144
145
  # pylint: disable=unused-argument
  def input_fn(is_training, data_dir, batch_size, *args, **kwargs):
    """Returns dataset filled with random data."""
    # Synthetic input should be within [0, 255].
146
    inputs = tf.random.truncated_normal(
Toby Boyd's avatar
Toby Boyd committed
147
148
149
150
151
152
        [batch_size] + [height, width, num_channels],
        dtype=dtype,
        mean=127,
        stddev=60,
        name='synthetic_inputs')

153
    labels = tf.random.uniform(
Toby Boyd's avatar
Toby Boyd committed
154
155
156
157
158
159
        [batch_size],
        minval=0,
        maxval=num_classes - 1,
        dtype=tf.int32,
        name='synthetic_labels')
    data = tf.data.Dataset.from_tensors((inputs, labels)).repeat()
160
    data = data.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
Toby Boyd's avatar
Toby Boyd committed
161
    return data
162
163
164
165

  return input_fn


166
def image_bytes_serving_input_fn(image_shape, dtype=tf.float32):
167
168
169
170
171
  """Serving input fn for raw jpeg images."""

  def _preprocess_image(image_bytes):
    """Preprocess a single raw image."""
    # Bounding box around the whole image.
172
    bbox = tf.constant([0.0, 0.0, 1.0, 1.0], dtype=dtype, shape=[1, 1, 4])
173
174
175
176
177
    height, width, num_channels = image_shape
    image = imagenet_preprocessing.preprocess_image(
        image_bytes, bbox, height, width, num_channels, is_training=False)
    return image

178
  image_bytes_list = tf.compat.v1.placeholder(
179
180
      shape=[None], dtype=tf.string, name='input_tensor')
  images = tf.map_fn(
181
      _preprocess_image, image_bytes_list, back_prop=False, dtype=dtype)
182
183
184
185
  return tf.estimator.export.TensorServingInputReceiver(
      images, {'image_bytes': image_bytes_list})


Toby Boyd's avatar
Toby Boyd committed
186
def override_flags_and_set_envars_for_gpu_thread_pool(flags_obj):
Toby Boyd's avatar
Toby Boyd committed
187
  """Override flags and set env_vars for performance.
Toby Boyd's avatar
Toby Boyd committed
188
189
190
191
192
193
194

  These settings exist to test the difference between using stock settings
  and manual tuning. It also shows some of the ENV_VARS that can be tweaked to
  squeeze a few extra examples per second.  These settings are defaulted to the
  current platform of interest, which changes over time.

  On systems with small numbers of cpu cores, e.g. under 8 logical cores,
Toby Boyd's avatar
Toby Boyd committed
195
196
  setting up a gpu thread pool with `tf_gpu_thread_mode=gpu_private` may perform
  poorly.
Toby Boyd's avatar
Toby Boyd committed
197
198
199
200
201

  Args:
    flags_obj: Current flags, which will be adjusted possibly overriding
    what has been set by the user on the command-line.
  """
Toby Boyd's avatar
Toby Boyd committed
202
  cpu_count = multiprocessing.cpu_count()
203
  tf.compat.v1.logging.info('Logical CPU cores: %s', cpu_count)
Toby Boyd's avatar
Toby Boyd committed
204
205
206
207
208
209

  # Sets up thread pool for each GPU for op scheduling.
  per_gpu_thread_count = 1
  total_gpu_thread_count = per_gpu_thread_count * flags_obj.num_gpus
  os.environ['TF_GPU_THREAD_MODE'] = flags_obj.tf_gpu_thread_mode
  os.environ['TF_GPU_THREAD_COUNT'] = str(per_gpu_thread_count)
210
211
212
213
  tf.compat.v1.logging.info('TF_GPU_THREAD_COUNT: %s',
                            os.environ['TF_GPU_THREAD_COUNT'])
  tf.compat.v1.logging.info('TF_GPU_THREAD_MODE: %s',
                            os.environ['TF_GPU_THREAD_MODE'])
Toby Boyd's avatar
Toby Boyd committed
214
215
216
217
218
219
220
221
222

  # Reduces general thread pool by number of threads used for GPU pool.
  main_thread_count = cpu_count - total_gpu_thread_count
  flags_obj.inter_op_parallelism_threads = main_thread_count

  # Sets thread count for tf.data. Logical cores minus threads assign to the
  # private GPU pool along with 2 thread per GPU for event monitoring and
  # sending / receiving tensors.
  num_monitoring_threads = 2 * flags_obj.num_gpus
Toby Boyd's avatar
Toby Boyd committed
223
224
  flags_obj.datasets_num_private_threads = (cpu_count - total_gpu_thread_count
                                            - num_monitoring_threads)
Toby Boyd's avatar
Toby Boyd committed
225
226


227
228
229
230
################################################################################
# Functions for running training/eval/validation loops for the model.
################################################################################
def learning_rate_with_decay(
231
232
    batch_size, batch_denom, num_images, boundary_epochs, decay_rates,
    base_lr=0.1, warmup=False):
233
234
235
236
237
238
239
240
241
242
243
  """Get a learning rate that decays step-wise as training progresses.

  Args:
    batch_size: the number of examples processed in each training batch.
    batch_denom: this value will be used to scale the base learning rate.
      `0.1 * batch size` is divided by this number, such that when
      batch_denom == batch_size, the initial learning rate will be 0.1.
    num_images: total number of images that will be used for training.
    boundary_epochs: list of ints representing the epochs at which we
      decay the learning rate.
    decay_rates: list of floats representing the decay rates to be used
244
245
      for scaling the learning rate. It should have one more element
      than `boundary_epochs`, and all elements should have the same type.
246
247
    base_lr: Initial learning rate scaled based on batch_denom.
    warmup: Run a 5 epoch warmup to the initial lr.
248
249
250
251
252
  Returns:
    Returns a function that takes a single argument - the number of batches
    trained so far (global_step)- and returns the learning rate to be used
    for training the next batch.
  """
253
  initial_learning_rate = base_lr * batch_size / batch_denom
254
255
  batches_per_epoch = num_images / batch_size

Taylor Robie's avatar
Taylor Robie committed
256
257
258
  # Reduce the learning rate at certain epochs.
  # CIFAR-10: divide by 10 at epoch 100, 150, and 200
  # ImageNet: divide by 10 at epoch 30, 60, 80, and 90
259
260
261
262
  boundaries = [int(batches_per_epoch * epoch) for epoch in boundary_epochs]
  vals = [initial_learning_rate * decay for decay in decay_rates]

  def learning_rate_fn(global_step):
263
    """Builds scaled learning rate function with 5 epoch warm up."""
264
    lr = tf.compat.v1.train.piecewise_constant(global_step, boundaries, vals)
265
266
267
268
269
    if warmup:
      warmup_steps = int(batches_per_epoch * 5)
      warmup_lr = (
          initial_learning_rate * tf.cast(global_step, tf.float32) / tf.cast(
              warmup_steps, tf.float32))
270
271
272
      return tf.cond(pred=global_step < warmup_steps,
                     true_fn=lambda: warmup_lr,
                     false_fn=lambda: lr)
273
    return lr
274

pkanwar23's avatar
pkanwar23 committed
275
276
277
278
279
280
281
282
283
  def poly_rate_fn(global_step):
    """Handles linear scaling rule, gradual warmup, and LR decay.

    The learning rate starts at 0, then it increases linearly per step.  After
    FLAGS.poly_warmup_epochs, we reach the base learning rate (scaled to account
    for batch size). The learning rate is then decayed using a polynomial rate
    decay schedule with power 2.0.

    Args:
Toby Boyd's avatar
Toby Boyd committed
284
      global_step: the current global_step
pkanwar23's avatar
pkanwar23 committed
285
286

    Returns:
Toby Boyd's avatar
Toby Boyd committed
287
      returns the current learning rate
pkanwar23's avatar
pkanwar23 committed
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
    """

    # Learning rate schedule for LARS polynomial schedule
    if flags.FLAGS.batch_size < 8192:
      plr = 5.0
      w_epochs = 5
    elif flags.FLAGS.batch_size < 16384:
      plr = 10.0
      w_epochs = 5
    elif flags.FLAGS.batch_size < 32768:
      plr = 25.0
      w_epochs = 5
    else:
      plr = 32.0
      w_epochs = 14

    w_steps = int(w_epochs * batches_per_epoch)
    wrate = (plr * tf.cast(global_step, tf.float32) / tf.cast(
        w_steps, tf.float32))

    # TODO(pkanwar): use a flag to help calc num_epochs.
    num_epochs = 90
    train_steps = batches_per_epoch * num_epochs

    min_step = tf.constant(1, dtype=tf.int64)
    decay_steps = tf.maximum(min_step, tf.subtract(global_step, w_steps))
    poly_rate = tf.train.polynomial_decay(
        plr,
        decay_steps,
        train_steps - w_steps + 1,
        power=2.0)
    return tf.where(global_step <= w_steps, wrate, poly_rate)

  # For LARS we have a new learning rate schedule
  if flags.FLAGS.enable_lars:
    return poly_rate_fn

325
326
327
328
329
  return learning_rate_fn


def resnet_model_fn(features, labels, mode, model_class,
                    resnet_size, weight_decay, learning_rate_fn, momentum,
330
                    data_format, resnet_version, loss_scale,
Zac Wellmer's avatar
Zac Wellmer committed
331
                    loss_filter_fn=None, dtype=resnet_model.DEFAULT_DTYPE,
pkanwar23's avatar
pkanwar23 committed
332
                    fine_tune=False, label_smoothing=0.0):
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
  """Shared functionality for different resnet model_fns.

  Initializes the ResnetModel representing the model layers
  and uses that model to build the necessary EstimatorSpecs for
  the `mode` in question. For training, this means building losses,
  the optimizer, and the train op that get passed into the EstimatorSpec.
  For evaluation and prediction, the EstimatorSpec is returned without
  a train op, but with the necessary parameters for the given mode.

  Args:
    features: tensor representing input images
    labels: tensor representing class labels for all input images
    mode: current estimator mode; should be one of
      `tf.estimator.ModeKeys.TRAIN`, `EVALUATE`, `PREDICT`
    model_class: a class representing a TensorFlow model that has a __call__
      function. We assume here that this is a subclass of ResnetModel.
    resnet_size: A single integer for the size of the ResNet model.
    weight_decay: weight decay loss rate used to regularize learned variables.
    learning_rate_fn: function that returns the current learning rate given
      the current global_step
    momentum: momentum term used for optimization
    data_format: Input format ('channels_last', 'channels_first', or None).
      If set to None, the format is dependent on whether a GPU is available.
356
357
    resnet_version: Integer representing which version of the ResNet network to
      use. See README for details. Valid values: [1, 2]
358
359
    loss_scale: The factor to scale the loss for numerical stability. A detailed
      summary is present in the arg parser help text.
360
361
362
363
    loss_filter_fn: function that takes a string variable name and returns
      True if the var should be included in loss calculation, and False
      otherwise. If None, batch_normalization variables will be excluded
      from the loss.
364
    dtype: the TensorFlow dtype to use for calculations.
Zac Wellmer's avatar
Zac Wellmer committed
365
    fine_tune: If True only train the dense layers(final layers).
Toby Boyd's avatar
Toby Boyd committed
366
    label_smoothing: If greater than 0 then smooth the labels.
367
368
369
370
371
372
373

  Returns:
    EstimatorSpec parameterized according to the input params and the
    current mode.
  """

  # Generate a summary node for the images
374
  tf.compat.v1.summary.image('images', features, max_outputs=6)
375
376
  # Checks that features/images have same data type being used for calculations.
  assert features.dtype == dtype
377

378
379
  model = model_class(resnet_size, data_format, resnet_version=resnet_version,
                      dtype=dtype)
380

381
382
  logits = model(features, mode == tf.estimator.ModeKeys.TRAIN)

383
384
385
386
387
  # This acts as a no-op if the logits are already in fp32 (provided logits are
  # not a SparseTensor). If dtype is is low precision, logits must be cast to
  # fp32 for numerical stability.
  logits = tf.cast(logits, tf.float32)

388
  predictions = {
389
      'classes': tf.argmax(input=logits, axis=1),
390
391
392
393
      'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
  }

  if mode == tf.estimator.ModeKeys.PREDICT:
394
395
396
397
398
399
400
    # Return the predictions and the specification for serving a SavedModel
    return tf.estimator.EstimatorSpec(
        mode=mode,
        predictions=predictions,
        export_outputs={
            'predict': tf.estimator.export.PredictOutput(predictions)
        })
401
402

  # Calculate loss, which includes softmax cross entropy and L2 regularization.
pkanwar23's avatar
pkanwar23 committed
403
404
405
406
407
408
  if label_smoothing != 0.0:
    one_hot_labels = tf.one_hot(labels, 1001)
    cross_entropy = tf.losses.softmax_cross_entropy(
        logits=logits, onehot_labels=one_hot_labels,
        label_smoothing=label_smoothing)
  else:
Toby Boyd's avatar
Toby Boyd committed
409
    cross_entropy = tf.compat.v1.losses.sparse_softmax_cross_entropy(
pkanwar23's avatar
pkanwar23 committed
410
        logits=logits, labels=labels)
411
412
413

  # Create a tensor named cross_entropy for logging purposes.
  tf.identity(cross_entropy, name='cross_entropy')
414
  tf.compat.v1.summary.scalar('cross_entropy', cross_entropy)
415
416
417

  # If no loss_filter_fn is passed, assume we want the default behavior,
  # which is that batch_normalization variables are excluded from loss.
Karmel Allison's avatar
Karmel Allison committed
418
419
420
  def exclude_batch_norm(name):
    return 'batch_normalization' not in name
  loss_filter_fn = loss_filter_fn or exclude_batch_norm
421

422
423
424
  # Add weight decay to the loss. We need to scale the regularization loss
  # manually as losses other than in tf.losses and tf.keras.losses don't scale
  # automatically.
425
  l2_loss = weight_decay * tf.add_n(
426
      # loss is computed using fp32 for numerical stability.
427
428
      [
          tf.nn.l2_loss(tf.cast(v, tf.float32))
429
          for v in tf.compat.v1.trainable_variables()
430
431
          if loss_filter_fn(v.name)
      ]) / tf.distribute.get_strategy().num_replicas_in_sync
432
  tf.compat.v1.summary.scalar('l2_loss', l2_loss)
433
  loss = cross_entropy + l2_loss
434
435

  if mode == tf.estimator.ModeKeys.TRAIN:
436
    global_step = tf.compat.v1.train.get_or_create_global_step()
437
438
439
440
441

    learning_rate = learning_rate_fn(global_step)

    # Create a tensor named learning_rate for logging purposes
    tf.identity(learning_rate, name='learning_rate')
442
    tf.compat.v1.summary.scalar('learning_rate', learning_rate)
443

pkanwar23's avatar
pkanwar23 committed
444
445
446
447
448
449
450
451
452
453
454
    if flags.FLAGS.enable_lars:
      optimizer = tf.contrib.opt.LARSOptimizer(
          learning_rate,
          momentum=momentum,
          weight_decay=weight_decay,
          skip_list=['batch_normalization', 'bias'])
    else:
      optimizer = tf.compat.v1.train.MomentumOptimizer(
          learning_rate=learning_rate,
          momentum=momentum
      )
455

Zac Wellmer's avatar
Zac Wellmer committed
456
    def _dense_grad_filter(gvs):
457
458
459
460
      """Only apply gradient updates to the final layer.

      This function is used for fine tuning.

Zac Wellmer's avatar
Zac Wellmer committed
461
      Args:
462
        gvs: list of tuples with gradients and variable info
Zac Wellmer's avatar
Zac Wellmer committed
463
      Returns:
464
465
        filtered gradients so that only the dense layer remains
      """
Zac Wellmer's avatar
Zac Wellmer committed
466
467
      return [(g, v) for g, v in gvs if 'dense' in v.name]

468
469
470
471
472
473
    if loss_scale != 1:
      # When computing fp16 gradients, often intermediate tensor values are
      # so small, they underflow to 0. To avoid this, we multiply the loss by
      # loss_scale to make these tensor values loss_scale times bigger.
      scaled_grad_vars = optimizer.compute_gradients(loss * loss_scale)

Zac Wellmer's avatar
Zac Wellmer committed
474
475
476
      if fine_tune:
        scaled_grad_vars = _dense_grad_filter(scaled_grad_vars)

477
478
479
480
481
482
      # Once the gradient computation is complete we can scale the gradients
      # back to the correct scale before passing them to the optimizer.
      unscaled_grad_vars = [(grad / loss_scale, var)
                            for grad, var in scaled_grad_vars]
      minimize_op = optimizer.apply_gradients(unscaled_grad_vars, global_step)
    else:
Zac Wellmer's avatar
Zac Wellmer committed
483
484
485
486
      grad_vars = optimizer.compute_gradients(loss)
      if fine_tune:
        grad_vars = _dense_grad_filter(grad_vars)
      minimize_op = optimizer.apply_gradients(grad_vars, global_step)
487

488
    update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
489
    train_op = tf.group(minimize_op, update_ops)
490
491
492
  else:
    train_op = None

493
494
495
  accuracy = tf.compat.v1.metrics.accuracy(labels, predictions['classes'])
  accuracy_top_5 = tf.compat.v1.metrics.mean(
      tf.nn.in_top_k(predictions=logits, targets=labels, k=5, name='top_5_op'))
496
497
  metrics = {'accuracy': accuracy,
             'accuracy_top_5': accuracy_top_5}
498
499
500

  # Create a tensor named train_accuracy for logging purposes
  tf.identity(accuracy[1], name='train_accuracy')
501
  tf.identity(accuracy_top_5[1], name='train_accuracy_top_5')
502
503
  tf.compat.v1.summary.scalar('train_accuracy', accuracy[1])
  tf.compat.v1.summary.scalar('train_accuracy_top_5', accuracy_top_5[1])
504
505
506
507
508

  return tf.estimator.EstimatorSpec(
      mode=mode,
      predictions=predictions,
      loss=loss,
509
510
      train_op=train_op,
      eval_metric_ops=metrics)
511
512


513
514
def resnet_main(
    flags_obj, model_function, input_function, dataset_name, shape=None):
515
516
517
  """Shared main loop for ResNet Models.

  Args:
518
519
    flags_obj: An object containing parsed flags. See define_resnet_flags()
      for details.
520
521
522
523
524
    model_function: the function that instantiates the Model and builds the
      ops for train/eval. This will be passed directly into the estimator.
    input_function: the function that processes the dataset and returns a
      dataset that the estimator can train on. This will be wrapped with
      all the relevant flags for running and passed to estimator.
525
526
    dataset_name: the name of the dataset for training and evaluation. This is
      used for logging purpose.
527
    shape: list of ints representing the shape of the images used for training.
528
      This is only used if flags_obj.export_dir is passed.
529
530
531

  Returns:
    Dict of results of the run.
532
  """
Karmel Allison's avatar
Karmel Allison committed
533

534
535
  model_helpers.apply_clean(flags.FLAGS)

Toby Boyd's avatar
Toby Boyd committed
536
  # Ensures flag override logic is only executed if explicitly triggered.
Toby Boyd's avatar
Toby Boyd committed
537
  if flags_obj.tf_gpu_thread_mode:
Toby Boyd's avatar
Toby Boyd committed
538
    override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
Toby Boyd's avatar
Toby Boyd committed
539

540
541
542
543
  # Configures cluster spec for distribution strategy.
  num_workers = distribution_utils.configure_cluster(flags_obj.worker_hosts,
                                                     flags_obj.task_index)

Toby Boyd's avatar
Toby Boyd committed
544
545
  # Creates session config. allow_soft_placement = True, is required for
  # multi-GPU and is not harmful for other modes.
546
  session_config = tf.compat.v1.ConfigProto(
Toby Boyd's avatar
Toby Boyd committed
547
548
549
      inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
      intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
      allow_soft_placement=True)
550

551
  distribution_strategy = distribution_utils.get_distribution_strategy(
552
553
      distribution_strategy=flags_obj.distribution_strategy,
      num_gpus=flags_core.get_num_gpus(flags_obj),
554
      num_workers=num_workers,
555
      all_reduce_alg=flags_obj.all_reduce_alg)
556

Toby Boyd's avatar
Toby Boyd committed
557
  # Creates a `RunConfig` that checkpoints every 24 hours which essentially
Toby Boyd's avatar
Toby Boyd committed
558
  # results in checkpoints determined only by `epochs_between_evals`.
559
  run_config = tf.estimator.RunConfig(
Toby Boyd's avatar
Toby Boyd committed
560
561
      train_distribute=distribution_strategy,
      session_config=session_config,
562
      save_checkpoints_secs=60*60*24,
563
      save_checkpoints_steps=None)
564

Toby Boyd's avatar
Toby Boyd committed
565
  # Initializes model with all but the dense layer from pretrained ResNet.
Zac Wellmer's avatar
Zac Wellmer committed
566
567
568
569
570
571
572
  if flags_obj.pretrained_model_checkpoint_path is not None:
    warm_start_settings = tf.estimator.WarmStartSettings(
        flags_obj.pretrained_model_checkpoint_path,
        vars_to_warm_start='^(?!.*dense)')
  else:
    warm_start_settings = None

573
  classifier = tf.estimator.Estimator(
574
      model_fn=model_function, model_dir=flags_obj.model_dir, config=run_config,
Zac Wellmer's avatar
Zac Wellmer committed
575
      warm_start_from=warm_start_settings, params={
576
577
578
          'resnet_size': int(flags_obj.resnet_size),
          'data_format': flags_obj.data_format,
          'batch_size': flags_obj.batch_size,
579
          'resnet_version': int(flags_obj.resnet_version),
580
          'loss_scale': flags_core.get_loss_scale(flags_obj),
Zac Wellmer's avatar
Zac Wellmer committed
581
          'dtype': flags_core.get_tf_dtype(flags_obj),
582
583
          'fine_tune': flags_obj.fine_tune,
          'num_workers': num_workers,
584
585
      })

586
587
588
589
  run_params = {
      'batch_size': flags_obj.batch_size,
      'dtype': flags_core.get_tf_dtype(flags_obj),
      'resnet_size': flags_obj.resnet_size,
590
      'resnet_version': flags_obj.resnet_version,
591
592
      'synthetic_data': flags_obj.use_synthetic_data,
      'train_epochs': flags_obj.train_epochs,
593
      'num_workers': num_workers,
594
  }
595
  if flags_obj.use_synthetic_data:
596
    dataset_name = dataset_name + '-synthetic'
597

598
  benchmark_logger = logger.get_benchmark_logger()
599
600
  benchmark_logger.log_run_info('resnet', dataset_name, run_params,
                                test_id=flags_obj.benchmark_test_id)
601

602
  train_hooks = hooks_helper.get_train_hooks(
603
      flags_obj.hooks,
604
      model_dir=flags_obj.model_dir,
605
      batch_size=flags_obj.batch_size)
606

607
  def input_fn_train(num_epochs, input_context=None):
608
    return input_function(
Toby Boyd's avatar
Toby Boyd committed
609
610
        is_training=True,
        data_dir=flags_obj.data_dir,
611
        batch_size=distribution_utils.per_device_batch_size(
612
            flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
Taylor Robie's avatar
Taylor Robie committed
613
        num_epochs=num_epochs,
Toby Boyd's avatar
Toby Boyd committed
614
615
        dtype=flags_core.get_tf_dtype(flags_obj),
        datasets_num_private_threads=flags_obj.datasets_num_private_threads,
616
617
        num_parallel_batches=flags_obj.datasets_num_parallel_batches,
        input_context=input_context)
618

619
  def input_fn_eval():
620
    return input_function(
Toby Boyd's avatar
Toby Boyd committed
621
622
        is_training=False,
        data_dir=flags_obj.data_dir,
623
        batch_size=distribution_utils.per_device_batch_size(
624
            flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
625
626
        num_epochs=1,
        dtype=flags_core.get_tf_dtype(flags_obj))
Taylor Robie's avatar
Taylor Robie committed
627

628
629
630
  train_epochs = (0 if flags_obj.eval_only or not flags_obj.train_epochs else
                  flags_obj.train_epochs)

631
  use_train_and_evaluate = flags_obj.use_train_and_evaluate or (
632
633
      distribution_strategy.__class__.__name__ in [
          'CollectiveAllReduceStrategy', 'MultiWorkerMirroredStrategy'])
634
635
  if use_train_and_evaluate:
    train_spec = tf.estimator.TrainSpec(
636
637
638
        input_fn=lambda input_context=None: input_fn_train(
            train_epochs, input_context=input_context),
        hooks=train_hooks,
639
640
641
642
643
644
        max_steps=flags_obj.max_train_steps)
    eval_spec = tf.estimator.EvalSpec(input_fn=input_fn_eval,
                                      steps=flags_obj.max_train_steps)
    tf.compat.v1.logging.info('Starting to train and evaluate.')
    eval_results, _ = tf.estimator.train_and_evaluate(classifier, train_spec,
                                                      eval_spec)
Qianli Scott Zhu's avatar
Qianli Scott Zhu committed
645
    benchmark_logger.log_evaluation_result(eval_results)
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
  else:
    if train_epochs == 0:
      # If --eval_only is set, perform a single loop with zero train epochs.
      schedule, n_loops = [0], 1
    else:
      # Compute the number of times to loop while training. All but the last
      # pass will train for `epochs_between_evals` epochs, while the last will
      # train for the number needed to reach `training_epochs`. For instance if
      #   train_epochs = 25 and epochs_between_evals = 10
      # schedule will be set to [10, 10, 5]. That is to say, the loop will:
      #   Train for 10 epochs and then evaluate.
      #   Train for another 10 epochs and then evaluate.
      #   Train for a final 5 epochs (to reach 25 epochs) and then evaluate.
      n_loops = math.ceil(train_epochs / flags_obj.epochs_between_evals)
      schedule = [flags_obj.epochs_between_evals for _ in range(int(n_loops))]
      schedule[-1] = train_epochs - sum(schedule[:-1])  # over counting.

    for cycle_index, num_train_epochs in enumerate(schedule):
      tf.compat.v1.logging.info('Starting cycle: %d/%d', cycle_index,
                                int(n_loops))

      if num_train_epochs:
668
669
670
671
        # Since we are calling classifier.train immediately in each loop, the
        # value of num_train_epochs in the lambda function will not be changed
        # before it is used. So it is safe to ignore the pylint error here
        # pylint: disable=cell-var-from-loop
672
673
674
675
676
        classifier.train(
            input_fn=lambda input_context=None: input_fn_train(
                num_train_epochs, input_context=input_context),
            hooks=train_hooks,
            max_steps=flags_obj.max_train_steps)
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692

      # flags_obj.max_train_steps is generally associated with testing and
      # profiling. As a result it is frequently called with synthetic data,
      # which will iterate forever. Passing steps=flags_obj.max_train_steps
      # allows the eval (which is generally unimportant in those circumstances)
      # to terminate.  Note that eval will run for max_train_steps each loop,
      # regardless of the global_step count.
      tf.compat.v1.logging.info('Starting to evaluate.')
      eval_results = classifier.evaluate(input_fn=input_fn_eval,
                                         steps=flags_obj.max_train_steps)

      benchmark_logger.log_evaluation_result(eval_results)

      if model_helpers.past_stop_threshold(
          flags_obj.stop_threshold, eval_results['accuracy']):
        break
693

694
  if flags_obj.export_dir is not None:
695
    # Exports a saved model for the given classifier.
696
    export_dtype = flags_core.get_tf_dtype(flags_obj)
697
    if flags_obj.image_bytes_as_serving_input:
698
699
      input_receiver_fn = functools.partial(
          image_bytes_serving_input_fn, shape, dtype=export_dtype)
700
701
    else:
      input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
702
703
704
          shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
    classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn,
                                 strip_default_attrs=True)
705
706
707
708
709
710
711

  stats = {}
  stats['eval_results'] = eval_results
  stats['train_hooks'] = train_hooks

  return stats

712

713
def define_resnet_flags(resnet_size_choices=None, dynamic_loss_scale=False):
714
715
  """Add flags and validators for ResNet."""
  flags_core.define_base()
Toby Boyd's avatar
Toby Boyd committed
716
717
718
  flags_core.define_performance(num_parallel_calls=False,
                                tf_gpu_thread_mode=True,
                                datasets_num_private_threads=True,
719
720
                                datasets_num_parallel_batches=True,
                                dynamic_loss_scale=dynamic_loss_scale)
721
722
723
  flags_core.define_image()
  flags_core.define_benchmark()
  flags.adopt_module_key_flags(flags_core)
724

725
  flags.DEFINE_enum(
Toby Boyd's avatar
Toby Boyd committed
726
      name='resnet_version', short_name='rv', default='1',
727
      enum_values=['1', '2'],
728
729
      help=flags_core.help_wrap(
          'Version of ResNet. (1 or 2) See README.md for details.'))
Zac Wellmer's avatar
Zac Wellmer committed
730
731
732
733
734
735
736
737
738
  flags.DEFINE_bool(
      name='fine_tune', short_name='ft', default=False,
      help=flags_core.help_wrap(
          'If True do not train any parameters except for the final layer.'))
  flags.DEFINE_string(
      name='pretrained_model_checkpoint_path', short_name='pmcp', default=None,
      help=flags_core.help_wrap(
          'If not None initialize all the network except the final layer with '
          'these values'))
Taylor Robie's avatar
Taylor Robie committed
739
  flags.DEFINE_boolean(
740
      name='eval_only', default=False,
Taylor Robie's avatar
Taylor Robie committed
741
742
      help=flags_core.help_wrap('Skip training and only perform evaluation on '
                                'the latest checkpoint.'))
743
  flags.DEFINE_boolean(
Toby Boyd's avatar
Toby Boyd committed
744
      name='image_bytes_as_serving_input', default=False,
745
746
747
748
749
750
751
      help=flags_core.help_wrap(
          'If True exports savedmodel with serving signature that accepts '
          'JPEG image bytes instead of a fixed size [HxWxC] tensor that '
          'represents the image. The former is easier to use for serving at '
          'the expense of image resize/cropping being done as part of model '
          'inference. Note, this flag only applies to ImageNet and cannot '
          'be used for CIFAR.'))
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
  flags.DEFINE_boolean(
      name='use_train_and_evaluate', default=False,
      help=flags_core.help_wrap(
          'If True, uses `tf.estimator.train_and_evaluate` for the training '
          'and evaluation loop, instead of separate calls to `classifier.train '
          'and `classifier.evaluate`, which is the default behavior.'))
  flags.DEFINE_string(
      name='worker_hosts', default=None,
      help=flags_core.help_wrap(
          'Comma-separated list of worker ip:port pairs for running '
          'multi-worker models with DistributionStrategy.  The user would '
          'start the program on each host with identical value for this flag.'))
  flags.DEFINE_integer(
      name='task_index', default=-1,
      help=flags_core.help_wrap('If multi-worker training, the task_index of '
                                'this worker.'))
pkanwar23's avatar
pkanwar23 committed
768
769
770
771
772
773
774
775
776
777
778
779
780
  flags.DEFINE_bool(
      name='enable_lars', default=False,
      help=flags_core.help_wrap(
          'Enable LARS optimizer for large batch training.'))
  flags.DEFINE_float(
      name='label_smoothing', default=0.0,
      help=flags_core.help_wrap(
          'Label smoothing parameter used in the softmax_cross_entropy'))
  flags.DEFINE_float(
      name='weight_decay', default=1e-4,
      help=flags_core.help_wrap(
          'Weight decay coefficiant for l2 regularization.'))

781
782
783
  choice_kwargs = dict(
      name='resnet_size', short_name='rs', default='50',
      help=flags_core.help_wrap('The size of the ResNet model to use.'))
784

785
786
787
788
  if resnet_size_choices is None:
    flags.DEFINE_string(**choice_kwargs)
  else:
    flags.DEFINE_enum(enum_values=resnet_size_choices, **choice_kwargs)