resnet_run_loop.py 32.8 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
                           num_parallel_batches=1):
Karmel Allison's avatar
Karmel Allison committed
57
  """Given a Dataset with raw records, return an iterator over the records.
58
59
60
61
62
63
64
65
66
67
68

  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.
69
    dtype: Data type to use for images/features.
Toby Boyd's avatar
Toby Boyd committed
70
71
72
    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.
73
74
75
76

  Returns:
    Dataset of (image, label) pairs ready for iteration.
  """
77
78
79
80
81
82
83
84
  # 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)
85

Haoyu Zhang's avatar
Haoyu Zhang committed
86
87
88
89
90
  # 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)

91
92
  # Prefetches a batch at a time to smooth out the time taken to load input
  # files for shuffling and processing.
93
94
  dataset = dataset.prefetch(buffer_size=batch_size)
  if is_training:
95
    # Shuffles records before repeating to respect epoch boundaries.
96
97
    dataset = dataset.shuffle(buffer_size=shuffle_buffer)

98
  # Repeats the dataset for the number of epochs to train.
99
100
  dataset = dataset.repeat(num_epochs)

101
  # Parses the raw records into images and labels.
Haoyu Zhang's avatar
Haoyu Zhang committed
102
103
104
  dataset = dataset.map(
      lambda value: parse_record_fn(value, is_training, dtype),
      num_parallel_calls=tf.data.experimental.AUTOTUNE)
105
  dataset = dataset.batch(batch_size, drop_remainder=False)
106
107
108
109

  # 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
110
111
112
  # 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.
113
  dataset = dataset.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
114
115
116
117

  return dataset


Toby Boyd's avatar
Toby Boyd committed
118
119
120
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.
121

Toby Boyd's avatar
Toby Boyd committed
122
123
124
125
  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.
126
127
128
129
130
131
132

  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
133
    dtype: Data type for features/images.
134
135
136
137
138

  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
139
140
141
142
  # 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].
143
    inputs = tf.random.truncated_normal(
Toby Boyd's avatar
Toby Boyd committed
144
145
146
147
148
149
        [batch_size] + [height, width, num_channels],
        dtype=dtype,
        mean=127,
        stddev=60,
        name='synthetic_inputs')

150
    labels = tf.random.uniform(
Toby Boyd's avatar
Toby Boyd committed
151
152
153
154
155
156
        [batch_size],
        minval=0,
        maxval=num_classes - 1,
        dtype=tf.int32,
        name='synthetic_labels')
    data = tf.data.Dataset.from_tensors((inputs, labels)).repeat()
157
    data = data.prefetch(buffer_size=tf.data.experimental.AUTOTUNE)
Toby Boyd's avatar
Toby Boyd committed
158
    return data
159
160
161
162

  return input_fn


163
def image_bytes_serving_input_fn(image_shape, dtype=tf.float32):
164
165
166
167
168
  """Serving input fn for raw jpeg images."""

  def _preprocess_image(image_bytes):
    """Preprocess a single raw image."""
    # Bounding box around the whole image.
169
    bbox = tf.constant([0.0, 0.0, 1.0, 1.0], dtype=dtype, shape=[1, 1, 4])
170
171
172
173
174
    height, width, num_channels = image_shape
    image = imagenet_preprocessing.preprocess_image(
        image_bytes, bbox, height, width, num_channels, is_training=False)
    return image

175
  image_bytes_list = tf.compat.v1.placeholder(
176
177
      shape=[None], dtype=tf.string, name='input_tensor')
  images = tf.map_fn(
178
      _preprocess_image, image_bytes_list, back_prop=False, dtype=dtype)
179
180
181
182
  return tf.estimator.export.TensorServingInputReceiver(
      images, {'image_bytes': image_bytes_list})


Toby Boyd's avatar
Toby Boyd committed
183
def override_flags_and_set_envars_for_gpu_thread_pool(flags_obj):
Toby Boyd's avatar
Toby Boyd committed
184
  """Override flags and set env_vars for performance.
Toby Boyd's avatar
Toby Boyd committed
185
186
187
188
189
190
191

  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
192
193
  setting up a gpu thread pool with `tf_gpu_thread_mode=gpu_private` may perform
  poorly.
Toby Boyd's avatar
Toby Boyd committed
194
195
196
197
198

  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
199
  cpu_count = multiprocessing.cpu_count()
200
  tf.compat.v1.logging.info('Logical CPU cores: %s', cpu_count)
Toby Boyd's avatar
Toby Boyd committed
201
202
203
204
205
206

  # 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)
207
208
209
210
  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
211
212
213
214
215
216
217
218
219

  # 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
220
221
  flags_obj.datasets_num_private_threads = (cpu_count - total_gpu_thread_count
                                            - num_monitoring_threads)
Toby Boyd's avatar
Toby Boyd committed
222
223


224
225
226
227
################################################################################
# Functions for running training/eval/validation loops for the model.
################################################################################
def learning_rate_with_decay(
228
229
    batch_size, batch_denom, num_images, boundary_epochs, decay_rates,
    base_lr=0.1, warmup=False):
230
231
232
233
234
235
236
237
238
239
240
  """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
241
242
      for scaling the learning rate. It should have one more element
      than `boundary_epochs`, and all elements should have the same type.
243
244
    base_lr: Initial learning rate scaled based on batch_denom.
    warmup: Run a 5 epoch warmup to the initial lr.
245
246
247
248
249
  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.
  """
250
  initial_learning_rate = base_lr * batch_size / batch_denom
251
252
  batches_per_epoch = num_images / batch_size

Taylor Robie's avatar
Taylor Robie committed
253
254
255
  # 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
256
257
258
259
  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):
260
    """Builds scaled learning rate function with 5 epoch warm up."""
261
    lr = tf.compat.v1.train.piecewise_constant(global_step, boundaries, vals)
262
263
264
265
266
    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))
267
268
269
      return tf.cond(pred=global_step < warmup_steps,
                     true_fn=lambda: warmup_lr,
                     false_fn=lambda: lr)
270
    return lr
271

pkanwar23's avatar
pkanwar23 committed
272
273
274
275
276
277
278
279
280
  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
281
      global_step: the current global_step
pkanwar23's avatar
pkanwar23 committed
282
283

    Returns:
Toby Boyd's avatar
Toby Boyd committed
284
      returns the current learning rate
pkanwar23's avatar
pkanwar23 committed
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
    """

    # 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

322
323
324
325
326
  return learning_rate_fn


def resnet_model_fn(features, labels, mode, model_class,
                    resnet_size, weight_decay, learning_rate_fn, momentum,
327
                    data_format, resnet_version, loss_scale,
Zac Wellmer's avatar
Zac Wellmer committed
328
                    loss_filter_fn=None, dtype=resnet_model.DEFAULT_DTYPE,
pkanwar23's avatar
pkanwar23 committed
329
                    fine_tune=False, label_smoothing=0.0):
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
  """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.
353
354
    resnet_version: Integer representing which version of the ResNet network to
      use. See README for details. Valid values: [1, 2]
355
356
    loss_scale: The factor to scale the loss for numerical stability. A detailed
      summary is present in the arg parser help text.
357
358
359
360
    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.
361
    dtype: the TensorFlow dtype to use for calculations.
Zac Wellmer's avatar
Zac Wellmer committed
362
    fine_tune: If True only train the dense layers(final layers).
Toby Boyd's avatar
Toby Boyd committed
363
    label_smoothing: If greater than 0 then smooth the labels.
364
365
366
367
368
369
370

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

  # Generate a summary node for the images
371
  tf.compat.v1.summary.image('images', features, max_outputs=6)
372
373
  # Checks that features/images have same data type being used for calculations.
  assert features.dtype == dtype
374

375
376
  model = model_class(resnet_size, data_format, resnet_version=resnet_version,
                      dtype=dtype)
377

378
379
  logits = model(features, mode == tf.estimator.ModeKeys.TRAIN)

380
381
382
383
384
  # 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)

385
  predictions = {
386
      'classes': tf.argmax(input=logits, axis=1),
387
388
389
390
      'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
  }

  if mode == tf.estimator.ModeKeys.PREDICT:
391
392
393
394
395
396
397
    # 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)
        })
398
399

  # Calculate loss, which includes softmax cross entropy and L2 regularization.
pkanwar23's avatar
pkanwar23 committed
400
401
402
403
404
405
  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
406
    cross_entropy = tf.compat.v1.losses.sparse_softmax_cross_entropy(
pkanwar23's avatar
pkanwar23 committed
407
        logits=logits, labels=labels)
408
409
410

  # Create a tensor named cross_entropy for logging purposes.
  tf.identity(cross_entropy, name='cross_entropy')
411
  tf.compat.v1.summary.scalar('cross_entropy', cross_entropy)
412
413
414

  # 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
415
416
417
  def exclude_batch_norm(name):
    return 'batch_normalization' not in name
  loss_filter_fn = loss_filter_fn or exclude_batch_norm
418

419
420
421
  # 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.
422
  l2_loss = weight_decay * tf.add_n(
423
      # loss is computed using fp32 for numerical stability.
424
425
      [
          tf.nn.l2_loss(tf.cast(v, tf.float32))
426
          for v in tf.compat.v1.trainable_variables()
427
428
          if loss_filter_fn(v.name)
      ]) / tf.distribute.get_strategy().num_replicas_in_sync
429
  tf.compat.v1.summary.scalar('l2_loss', l2_loss)
430
  loss = cross_entropy + l2_loss
431
432

  if mode == tf.estimator.ModeKeys.TRAIN:
433
    global_step = tf.compat.v1.train.get_or_create_global_step()
434
435
436
437
438

    learning_rate = learning_rate_fn(global_step)

    # Create a tensor named learning_rate for logging purposes
    tf.identity(learning_rate, name='learning_rate')
439
    tf.compat.v1.summary.scalar('learning_rate', learning_rate)
440

pkanwar23's avatar
pkanwar23 committed
441
442
443
444
445
446
447
448
449
450
451
    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
      )
452

Zac Wellmer's avatar
Zac Wellmer committed
453
    def _dense_grad_filter(gvs):
454
455
456
457
      """Only apply gradient updates to the final layer.

      This function is used for fine tuning.

Zac Wellmer's avatar
Zac Wellmer committed
458
      Args:
459
        gvs: list of tuples with gradients and variable info
Zac Wellmer's avatar
Zac Wellmer committed
460
      Returns:
461
462
        filtered gradients so that only the dense layer remains
      """
Zac Wellmer's avatar
Zac Wellmer committed
463
464
      return [(g, v) for g, v in gvs if 'dense' in v.name]

465
466
467
468
469
470
    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
471
472
473
      if fine_tune:
        scaled_grad_vars = _dense_grad_filter(scaled_grad_vars)

474
475
476
477
478
479
      # 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
480
481
482
483
      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)
484

485
    update_ops = tf.compat.v1.get_collection(tf.compat.v1.GraphKeys.UPDATE_OPS)
486
    train_op = tf.group(minimize_op, update_ops)
487
488
489
  else:
    train_op = None

490
491
492
  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'))
493
494
  metrics = {'accuracy': accuracy,
             'accuracy_top_5': accuracy_top_5}
495
496
497

  # Create a tensor named train_accuracy for logging purposes
  tf.identity(accuracy[1], name='train_accuracy')
498
  tf.identity(accuracy_top_5[1], name='train_accuracy_top_5')
499
500
  tf.compat.v1.summary.scalar('train_accuracy', accuracy[1])
  tf.compat.v1.summary.scalar('train_accuracy_top_5', accuracy_top_5[1])
501
502
503
504
505

  return tf.estimator.EstimatorSpec(
      mode=mode,
      predictions=predictions,
      loss=loss,
506
507
      train_op=train_op,
      eval_metric_ops=metrics)
508
509


510
511
def resnet_main(
    flags_obj, model_function, input_function, dataset_name, shape=None):
512
513
514
  """Shared main loop for ResNet Models.

  Args:
515
516
    flags_obj: An object containing parsed flags. See define_resnet_flags()
      for details.
517
518
519
520
521
    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.
522
523
    dataset_name: the name of the dataset for training and evaluation. This is
      used for logging purpose.
524
    shape: list of ints representing the shape of the images used for training.
525
      This is only used if flags_obj.export_dir is passed.
526
527
528

  Returns:
    Dict of results of the run.
529
  """
Karmel Allison's avatar
Karmel Allison committed
530

531
532
  model_helpers.apply_clean(flags.FLAGS)

Toby Boyd's avatar
Toby Boyd committed
533
  # Ensures flag override logic is only executed if explicitly triggered.
Toby Boyd's avatar
Toby Boyd committed
534
  if flags_obj.tf_gpu_thread_mode:
Toby Boyd's avatar
Toby Boyd committed
535
    override_flags_and_set_envars_for_gpu_thread_pool(flags_obj)
Toby Boyd's avatar
Toby Boyd committed
536

537
538
539
540
  # 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
541
542
  # Creates session config. allow_soft_placement = True, is required for
  # multi-GPU and is not harmful for other modes.
543
  session_config = tf.compat.v1.ConfigProto(
Toby Boyd's avatar
Toby Boyd committed
544
545
546
      inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
      intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
      allow_soft_placement=True)
547

548
  distribution_strategy = distribution_utils.get_distribution_strategy(
549
550
      distribution_strategy=flags_obj.distribution_strategy,
      num_gpus=flags_core.get_num_gpus(flags_obj),
551
      num_workers=num_workers,
552
      all_reduce_alg=flags_obj.all_reduce_alg)
553

Toby Boyd's avatar
Toby Boyd committed
554
  # Creates a `RunConfig` that checkpoints every 24 hours which essentially
Toby Boyd's avatar
Toby Boyd committed
555
  # results in checkpoints determined only by `epochs_between_evals`.
556
  run_config = tf.estimator.RunConfig(
Toby Boyd's avatar
Toby Boyd committed
557
558
      train_distribute=distribution_strategy,
      session_config=session_config,
559
      save_checkpoints_secs=60*60*24,
560
      save_checkpoints_steps=None)
561

Toby Boyd's avatar
Toby Boyd committed
562
  # Initializes model with all but the dense layer from pretrained ResNet.
Zac Wellmer's avatar
Zac Wellmer committed
563
564
565
566
567
568
569
  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

570
  classifier = tf.estimator.Estimator(
571
      model_fn=model_function, model_dir=flags_obj.model_dir, config=run_config,
Zac Wellmer's avatar
Zac Wellmer committed
572
      warm_start_from=warm_start_settings, params={
573
574
575
          'resnet_size': int(flags_obj.resnet_size),
          'data_format': flags_obj.data_format,
          'batch_size': flags_obj.batch_size,
576
          'resnet_version': int(flags_obj.resnet_version),
577
          'loss_scale': flags_core.get_loss_scale(flags_obj),
Zac Wellmer's avatar
Zac Wellmer committed
578
          'dtype': flags_core.get_tf_dtype(flags_obj),
579
580
          'fine_tune': flags_obj.fine_tune,
          'num_workers': num_workers,
581
582
      })

583
584
585
586
  run_params = {
      'batch_size': flags_obj.batch_size,
      'dtype': flags_core.get_tf_dtype(flags_obj),
      'resnet_size': flags_obj.resnet_size,
587
      'resnet_version': flags_obj.resnet_version,
588
589
      'synthetic_data': flags_obj.use_synthetic_data,
      'train_epochs': flags_obj.train_epochs,
590
      'num_workers': num_workers,
591
  }
592
  if flags_obj.use_synthetic_data:
593
    dataset_name = dataset_name + '-synthetic'
594

595
  benchmark_logger = logger.get_benchmark_logger()
596
597
  benchmark_logger.log_run_info('resnet', dataset_name, run_params,
                                test_id=flags_obj.benchmark_test_id)
598

599
  train_hooks = hooks_helper.get_train_hooks(
600
      flags_obj.hooks,
601
      model_dir=flags_obj.model_dir,
602
      batch_size=flags_obj.batch_size)
603

604
  def input_fn_train(num_epochs, input_context=None):
605
    return input_function(
Toby Boyd's avatar
Toby Boyd committed
606
607
        is_training=True,
        data_dir=flags_obj.data_dir,
608
        batch_size=distribution_utils.per_device_batch_size(
609
            flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
Taylor Robie's avatar
Taylor Robie committed
610
        num_epochs=num_epochs,
Toby Boyd's avatar
Toby Boyd committed
611
612
        dtype=flags_core.get_tf_dtype(flags_obj),
        datasets_num_private_threads=flags_obj.datasets_num_private_threads,
613
614
        num_parallel_batches=flags_obj.datasets_num_parallel_batches,
        input_context=input_context)
615

616
  def input_fn_eval():
617
    return input_function(
Toby Boyd's avatar
Toby Boyd committed
618
619
        is_training=False,
        data_dir=flags_obj.data_dir,
620
        batch_size=distribution_utils.per_device_batch_size(
621
            flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
622
623
        num_epochs=1,
        dtype=flags_core.get_tf_dtype(flags_obj))
Taylor Robie's avatar
Taylor Robie committed
624

625
626
627
  train_epochs = (0 if flags_obj.eval_only or not flags_obj.train_epochs else
                  flags_obj.train_epochs)

628
  use_train_and_evaluate = flags_obj.use_train_and_evaluate or (
629
630
      distribution_strategy.__class__.__name__ in [
          'CollectiveAllReduceStrategy', 'MultiWorkerMirroredStrategy'])
631
632
  if use_train_and_evaluate:
    train_spec = tf.estimator.TrainSpec(
633
634
635
        input_fn=lambda input_context=None: input_fn_train(
            train_epochs, input_context=input_context),
        hooks=train_hooks,
636
637
638
639
640
641
        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
642
    benchmark_logger.log_evaluation_result(eval_results)
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
  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:
665
666
667
668
        # 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
669
670
671
672
673
        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)
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689

      # 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
690

691
  if flags_obj.export_dir is not None:
692
    # Exports a saved model for the given classifier.
693
    export_dtype = flags_core.get_tf_dtype(flags_obj)
694
    if flags_obj.image_bytes_as_serving_input:
695
696
      input_receiver_fn = functools.partial(
          image_bytes_serving_input_fn, shape, dtype=export_dtype)
697
698
    else:
      input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
699
700
701
          shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
    classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn,
                                 strip_default_attrs=True)
702
703
704
705
706
707
708

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

  return stats

709

710
711
712
def define_resnet_flags(resnet_size_choices=None):
  """Add flags and validators for ResNet."""
  flags_core.define_base()
Toby Boyd's avatar
Toby Boyd committed
713
714
715
716
  flags_core.define_performance(num_parallel_calls=False,
                                tf_gpu_thread_mode=True,
                                datasets_num_private_threads=True,
                                datasets_num_parallel_batches=True)
717
718
719
  flags_core.define_image()
  flags_core.define_benchmark()
  flags.adopt_module_key_flags(flags_core)
720

721
  flags.DEFINE_enum(
Toby Boyd's avatar
Toby Boyd committed
722
      name='resnet_version', short_name='rv', default='1',
723
      enum_values=['1', '2'],
724
725
      help=flags_core.help_wrap(
          'Version of ResNet. (1 or 2) See README.md for details.'))
Zac Wellmer's avatar
Zac Wellmer committed
726
727
728
729
730
731
732
733
734
  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
735
  flags.DEFINE_boolean(
736
      name='eval_only', default=False,
Taylor Robie's avatar
Taylor Robie committed
737
738
      help=flags_core.help_wrap('Skip training and only perform evaluation on '
                                'the latest checkpoint.'))
739
  flags.DEFINE_boolean(
Toby Boyd's avatar
Toby Boyd committed
740
      name='image_bytes_as_serving_input', default=False,
741
742
743
744
745
746
747
      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.'))
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
  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
764
765
766
767
768
769
770
771
772
773
774
775
776
  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.'))

777
778
779
  choice_kwargs = dict(
      name='resnet_size', short_name='rs', default='50',
      help=flags_core.help_wrap('The size of the ResNet model to use.'))
780

781
782
783
784
  if resnet_size_choices is None:
    flags.DEFINE_string(**choice_kwargs)
  else:
    flags.DEFINE_enum(enum_values=resnet_size_choices, **choice_kwargs)