resnet_run_loop.py 27.1 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
28
29
import os

30
# pylint: disable=g-bad-import-order
31
from absl import flags
32
import tensorflow as tf
Toby Boyd's avatar
Toby Boyd committed
33
34
from tensorflow.contrib.data.python.ops import threadpool
import multiprocessing
35
36

from official.resnet import resnet_model
37
from official.utils.flags import core as flags_core
38
from official.utils.export import export
39
40
from official.utils.logs import hooks_helper
from official.utils.logs import logger
41
from official.resnet import imagenet_preprocessing
42
from official.utils.misc import distribution_utils
43
from official.utils.misc import model_helpers
44
45
46
47
48


################################################################################
# Functions for input processing.
################################################################################
Toby Boyd's avatar
Toby Boyd committed
49
50
51
52
53
54
55
56
57
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,
                           num_parallel_batches=1):
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
76
77

  Returns:
    Dataset of (image, label) pairs ready for iteration.
  """
78

79
80
  # Prefetches a batch at a time to smooth out the time taken to load input
  # files for shuffling and processing.
81
82
  dataset = dataset.prefetch(buffer_size=batch_size)
  if is_training:
83
    # Shuffles records before repeating to respect epoch boundaries.
84
85
    dataset = dataset.shuffle(buffer_size=shuffle_buffer)

86
  # Repeats the dataset for the number of epochs to train.
87
88
  dataset = dataset.repeat(num_epochs)

89
  # Parses the raw records into images and labels.
90
91
  dataset = dataset.apply(
      tf.contrib.data.map_and_batch(
92
          lambda value: parse_record_fn(value, is_training, dtype),
93
          batch_size=batch_size,
Toby Boyd's avatar
Toby Boyd committed
94
          num_parallel_calls=num_parallel_batches,
95
          drop_remainder=False))
96
97
98
99

  # 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
100
101
102
  # 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.
103
  dataset = dataset.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)
104

Toby Boyd's avatar
Toby Boyd committed
105
106
107
108
109
110
111
112
  # Defines a specific size thread pool for tf.data operations.
  if datasets_num_private_threads:
    dataset = threadpool.override_threadpool(
        dataset,
        threadpool.PrivateThreadPool(
            datasets_num_private_threads,
            display_name='input_pipeline_thread_pool'))

113
114
115
  return dataset


Toby Boyd's avatar
Toby Boyd committed
116
117
118
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.
119

Toby Boyd's avatar
Toby Boyd committed
120
121
122
123
  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.
124
125
126
127
128
129
130

  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
131
    dtype: Data type for features/images.
132
133
134
135
136

  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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
  # 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].
    inputs = tf.truncated_normal(
        [batch_size] + [height, width, num_channels],
        dtype=dtype,
        mean=127,
        stddev=60,
        name='synthetic_inputs')

    labels = tf.random_uniform(
        [batch_size],
        minval=0,
        maxval=num_classes - 1,
        dtype=tf.int32,
        name='synthetic_labels')
    data = tf.data.Dataset.from_tensors((inputs, labels)).repeat()
    data = data.prefetch(buffer_size=tf.contrib.data.AUTOTUNE)
    return data
157
158
159
160

  return input_fn


161
def image_bytes_serving_input_fn(image_shape, dtype=tf.float32):
162
163
164
165
166
  """Serving input fn for raw jpeg images."""

  def _preprocess_image(image_bytes):
    """Preprocess a single raw image."""
    # Bounding box around the whole image.
167
    bbox = tf.constant([0.0, 0.0, 1.0, 1.0], dtype=dtype, shape=[1, 1, 4])
168
169
170
171
172
173
174
175
    height, width, num_channels = image_shape
    image = imagenet_preprocessing.preprocess_image(
        image_bytes, bbox, height, width, num_channels, is_training=False)
    return image

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


Toby Boyd's avatar
Toby Boyd committed
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
def set_environment_vars(flags_obj):
  """Adjust flags and set env_vars for performance.

  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,
  setting up a private thread pool for GPU with `tf_gpu_thread_mode=gpu_private`
  may perform poorly.

  Args:
    flags_obj: Current flags, which will be adjusted possibly overriding
    what has been set by the user on the command-line.

  Returns:
    tf.ConfigProto: session_config proto to add to the session.
  """
  if flags_obj.tf_gpu_thread_mode in ['gpu_private']:
    cpu_count = multiprocessing.cpu_count()
    print('Logical CPU cores:', cpu_count)

    # 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)
    print('TF_GPU_THREAD_COUNT:', os.environ['TF_GPU_THREAD_COUNT'])

    # 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
    num_private_threads = (cpu_count - total_gpu_thread_count
                           - num_monitoring_threads)
    flags_obj.datasets_num_private_threads = num_private_threads

  print('inter_op_parallelism_threads:', flags_obj.inter_op_parallelism_threads)
  print('intra_op_parallelism_threads:', flags_obj.intra_op_parallelism_threads)
  print('datasets_num_private_threads:', flags_obj.datasets_num_private_threads)

  # Create session config based on values of inter_op_parallelism_threads and
  # intra_op_parallelism_threads. Note that we default to having
  # allow_soft_placement = True, which is required for multi-GPU and not
  # harmful for other modes.
  session_config = tf.ConfigProto(
      inter_op_parallelism_threads=flags_obj.inter_op_parallelism_threads,
      intra_op_parallelism_threads=flags_obj.intra_op_parallelism_threads,
      allow_soft_placement=True)
  return session_config


238
239
240
241
################################################################################
# Functions for running training/eval/validation loops for the model.
################################################################################
def learning_rate_with_decay(
242
243
    batch_size, batch_denom, num_images, boundary_epochs, decay_rates,
    base_lr=0.1, warmup=False):
244
245
246
247
248
249
250
251
252
253
254
  """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
255
256
      for scaling the learning rate. It should have one more element
      than `boundary_epochs`, and all elements should have the same type.
257
258
    base_lr: Initial learning rate scaled based on batch_denom.
    warmup: Run a 5 epoch warmup to the initial lr.
259
260
261
262
263
  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.
  """
264
  initial_learning_rate = base_lr * batch_size / batch_denom
265
266
  batches_per_epoch = num_images / batch_size

Taylor Robie's avatar
Taylor Robie committed
267
268
269
  # 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
270
271
272
273
  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):
274
275
276
277
278
279
280
281
282
    """Builds scaled learning rate function with 5 epoch warm up."""
    lr = tf.train.piecewise_constant(global_step, boundaries, vals)
    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))
      return tf.cond(global_step < warmup_steps, lambda: warmup_lr, lambda: lr)
    return lr
283
284
285
286
287
288

  return learning_rate_fn


def resnet_model_fn(features, labels, mode, model_class,
                    resnet_size, weight_decay, learning_rate_fn, momentum,
289
                    data_format, resnet_version, loss_scale,
Zac Wellmer's avatar
Zac Wellmer committed
290
291
                    loss_filter_fn=None, dtype=resnet_model.DEFAULT_DTYPE,
                    fine_tune=False):
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
  """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.
315
316
    resnet_version: Integer representing which version of the ResNet network to
      use. See README for details. Valid values: [1, 2]
317
318
    loss_scale: The factor to scale the loss for numerical stability. A detailed
      summary is present in the arg parser help text.
319
320
321
322
    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.
323
    dtype: the TensorFlow dtype to use for calculations.
Zac Wellmer's avatar
Zac Wellmer committed
324
    fine_tune: If True only train the dense layers(final layers).
325
326
327
328
329
330
331
332

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

  # Generate a summary node for the images
  tf.summary.image('images', features, max_outputs=6)
333
334
  # Checks that features/images have same data type being used for calculations.
  assert features.dtype == dtype
335

336
337
  model = model_class(resnet_size, data_format, resnet_version=resnet_version,
                      dtype=dtype)
338

339
340
  logits = model(features, mode == tf.estimator.ModeKeys.TRAIN)

341
342
343
344
345
  # 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)

346
347
348
349
350
351
  predictions = {
      'classes': tf.argmax(logits, axis=1),
      'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
  }

  if mode == tf.estimator.ModeKeys.PREDICT:
352
353
354
355
356
357
358
    # 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)
        })
359
360

  # Calculate loss, which includes softmax cross entropy and L2 regularization.
361
362
  cross_entropy = tf.losses.sparse_softmax_cross_entropy(
      logits=logits, labels=labels)
363
364
365
366
367
368
369

  # Create a tensor named cross_entropy for logging purposes.
  tf.identity(cross_entropy, name='cross_entropy')
  tf.summary.scalar('cross_entropy', cross_entropy)

  # 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
370
371
372
  def exclude_batch_norm(name):
    return 'batch_normalization' not in name
  loss_filter_fn = loss_filter_fn or exclude_batch_norm
373
374

  # Add weight decay to the loss.
375
  l2_loss = weight_decay * tf.add_n(
376
377
      # loss is computed using fp32 for numerical stability.
      [tf.nn.l2_loss(tf.cast(v, tf.float32)) for v in tf.trainable_variables()
378
       if loss_filter_fn(v.name)])
379
380
  tf.summary.scalar('l2_loss', l2_loss)
  loss = cross_entropy + l2_loss
381
382
383
384
385
386
387
388
389
390
391
392

  if mode == tf.estimator.ModeKeys.TRAIN:
    global_step = tf.train.get_or_create_global_step()

    learning_rate = learning_rate_fn(global_step)

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

    optimizer = tf.train.MomentumOptimizer(
        learning_rate=learning_rate,
393
394
        momentum=momentum
    )
395

Zac Wellmer's avatar
Zac Wellmer committed
396
    def _dense_grad_filter(gvs):
397
398
399
400
      """Only apply gradient updates to the final layer.

      This function is used for fine tuning.

Zac Wellmer's avatar
Zac Wellmer committed
401
      Args:
402
        gvs: list of tuples with gradients and variable info
Zac Wellmer's avatar
Zac Wellmer committed
403
      Returns:
404
405
        filtered gradients so that only the dense layer remains
      """
Zac Wellmer's avatar
Zac Wellmer committed
406
407
      return [(g, v) for g, v in gvs if 'dense' in v.name]

408
409
410
411
412
413
    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
414
415
416
      if fine_tune:
        scaled_grad_vars = _dense_grad_filter(scaled_grad_vars)

417
418
419
420
421
422
      # 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
423
424
425
426
      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)
427

428
    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
429
    train_op = tf.group(minimize_op, update_ops)
430
431
432
  else:
    train_op = None

433
  accuracy = tf.metrics.accuracy(labels, predictions['classes'])
434
435
436
437
438
439
  accuracy_top_5 = tf.metrics.mean(tf.nn.in_top_k(predictions=logits,
                                                  targets=labels,
                                                  k=5,
                                                  name='top_5_op'))
  metrics = {'accuracy': accuracy,
             'accuracy_top_5': accuracy_top_5}
440
441
442

  # Create a tensor named train_accuracy for logging purposes
  tf.identity(accuracy[1], name='train_accuracy')
443
  tf.identity(accuracy_top_5[1], name='train_accuracy_top_5')
444
  tf.summary.scalar('train_accuracy', accuracy[1])
445
  tf.summary.scalar('train_accuracy_top_5', accuracy_top_5[1])
446
447
448
449
450

  return tf.estimator.EstimatorSpec(
      mode=mode,
      predictions=predictions,
      loss=loss,
451
452
      train_op=train_op,
      eval_metric_ops=metrics)
453
454


455
456
def resnet_main(
    flags_obj, model_function, input_function, dataset_name, shape=None):
457
458
459
  """Shared main loop for ResNet Models.

  Args:
460
461
    flags_obj: An object containing parsed flags. See define_resnet_flags()
      for details.
462
463
464
465
466
    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.
467
468
    dataset_name: the name of the dataset for training and evaluation. This is
      used for logging purpose.
469
    shape: list of ints representing the shape of the images used for training.
470
      This is only used if flags_obj.export_dir is passed.
471
  """
Karmel Allison's avatar
Karmel Allison committed
472

473
474
  model_helpers.apply_clean(flags.FLAGS)

Toby Boyd's avatar
Toby Boyd committed
475
  session_config = set_environment_vars(flags_obj)
476

477
478
  distribution_strategy = distribution_utils.get_distribution_strategy(
      flags_core.get_num_gpus(flags_obj), flags_obj.all_reduce_alg)
479

Toby Boyd's avatar
Toby Boyd committed
480
481
482
  # Creates a `RunConfig` that checkpoints every 24 hours which essentially
  # results in checkpoints at the end of each training loop as determined by
  # `epochs_between_evals`.  Doing it more often is a needless small cost.
483
  run_config = tf.estimator.RunConfig(
Toby Boyd's avatar
Toby Boyd committed
484
485
486
      train_distribute=distribution_strategy,
      session_config=session_config,
      save_checkpoints_secs=60*60*24)
487

Zac Wellmer's avatar
Zac Wellmer committed
488
489
490
491
492
493
494
495
  # initialize our model with all but the dense layer from pretrained resnet
  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

496
  classifier = tf.estimator.Estimator(
497
      model_fn=model_function, model_dir=flags_obj.model_dir, config=run_config,
Zac Wellmer's avatar
Zac Wellmer committed
498
      warm_start_from=warm_start_settings, params={
499
500
501
          'resnet_size': int(flags_obj.resnet_size),
          'data_format': flags_obj.data_format,
          'batch_size': flags_obj.batch_size,
502
          'resnet_version': int(flags_obj.resnet_version),
503
          'loss_scale': flags_core.get_loss_scale(flags_obj),
Zac Wellmer's avatar
Zac Wellmer committed
504
505
          'dtype': flags_core.get_tf_dtype(flags_obj),
          'fine_tune': flags_obj.fine_tune
506
507
      })

508
509
510
511
  run_params = {
      'batch_size': flags_obj.batch_size,
      'dtype': flags_core.get_tf_dtype(flags_obj),
      'resnet_size': flags_obj.resnet_size,
512
      'resnet_version': flags_obj.resnet_version,
513
514
515
      'synthetic_data': flags_obj.use_synthetic_data,
      'train_epochs': flags_obj.train_epochs,
  }
516
  if flags_obj.use_synthetic_data:
517
    dataset_name = dataset_name + '-synthetic'
518

519
  benchmark_logger = logger.get_benchmark_logger()
520
521
  benchmark_logger.log_run_info('resnet', dataset_name, run_params,
                                test_id=flags_obj.benchmark_test_id)
522

523
  train_hooks = hooks_helper.get_train_hooks(
524
      flags_obj.hooks,
525
      model_dir=flags_obj.model_dir,
526
      batch_size=flags_obj.batch_size)
527

Taylor Robie's avatar
Taylor Robie committed
528
  def input_fn_train(num_epochs):
529
    return input_function(
Toby Boyd's avatar
Toby Boyd committed
530
531
        is_training=True,
        data_dir=flags_obj.data_dir,
532
        batch_size=distribution_utils.per_device_batch_size(
533
            flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
Taylor Robie's avatar
Taylor Robie committed
534
        num_epochs=num_epochs,
Toby Boyd's avatar
Toby Boyd committed
535
536
537
        dtype=flags_core.get_tf_dtype(flags_obj),
        datasets_num_private_threads=flags_obj.datasets_num_private_threads,
        num_parallel_batches=flags_obj.num_parallel_calls)
538

539
  def input_fn_eval():
540
    return input_function(
Toby Boyd's avatar
Toby Boyd committed
541
542
        is_training=False,
        data_dir=flags_obj.data_dir,
543
        batch_size=distribution_utils.per_device_batch_size(
544
            flags_obj.batch_size, flags_core.get_num_gpus(flags_obj)),
545
546
        num_epochs=1,
        dtype=flags_core.get_tf_dtype(flags_obj))
Taylor Robie's avatar
Taylor Robie committed
547

Taylor Robie's avatar
Taylor Robie committed
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
  if flags_obj.eval_only or not flags_obj.train_epochs:
    # 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(flags_obj.train_epochs / flags_obj.epochs_between_evals)
    schedule = [flags_obj.epochs_between_evals for _ in range(int(n_loops))]
    schedule[-1] = flags_obj.train_epochs - sum(schedule[:-1])  # over counting.

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

    if num_train_epochs:
      classifier.train(input_fn=lambda: input_fn_train(num_train_epochs),
                       hooks=train_hooks, max_steps=flags_obj.max_train_steps)
570

571
    tf.logging.info('Starting to evaluate.')
572
573
574
575
576

    # 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.
577
578
579
    # Note that eval will run for max_train_steps each loop, regardless of the
    # global_step count.
    eval_results = classifier.evaluate(input_fn=input_fn_eval,
580
                                       steps=flags_obj.max_train_steps)
581

Qianli Scott Zhu's avatar
Qianli Scott Zhu committed
582
    benchmark_logger.log_evaluation_result(eval_results)
583

584
    if model_helpers.past_stop_threshold(
585
        flags_obj.stop_threshold, eval_results['accuracy']):
586
587
      break

588
  if flags_obj.export_dir is not None:
589
    # Exports a saved model for the given classifier.
590
    export_dtype = flags_core.get_tf_dtype(flags_obj)
591
    if flags_obj.image_bytes_as_serving_input:
592
593
      input_receiver_fn = functools.partial(
          image_bytes_serving_input_fn, shape, dtype=export_dtype)
594
595
    else:
      input_receiver_fn = export.build_tensor_serving_input_receiver_fn(
596
597
598
          shape, batch_size=flags_obj.batch_size, dtype=export_dtype)
    classifier.export_savedmodel(flags_obj.export_dir, input_receiver_fn,
                                 strip_default_attrs=True)
599
600


601
602
603
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
604
  flags_core.define_performance()
605
606
607
  flags_core.define_image()
  flags_core.define_benchmark()
  flags.adopt_module_key_flags(flags_core)
608

609
  flags.DEFINE_enum(
Toby Boyd's avatar
Toby Boyd committed
610
      name='resnet_version', short_name='rv', default='1',
611
      enum_values=['1', '2'],
612
613
      help=flags_core.help_wrap(
          'Version of ResNet. (1 or 2) See README.md for details.'))
Zac Wellmer's avatar
Zac Wellmer committed
614
615
616
617
618
619
620
621
622
  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
623
  flags.DEFINE_boolean(
624
      name='eval_only', default=False,
Taylor Robie's avatar
Taylor Robie committed
625
626
      help=flags_core.help_wrap('Skip training and only perform evaluation on '
                                'the latest checkpoint.'))
627
  flags.DEFINE_boolean(
Toby Boyd's avatar
Toby Boyd committed
628
      name='image_bytes_as_serving_input', default=False,
629
630
631
632
633
634
635
      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.'))
636

637
638
639
  choice_kwargs = dict(
      name='resnet_size', short_name='rs', default='50',
      help=flags_core.help_wrap('The size of the ResNet model to use.'))
640

641
642
643
644
  if resnet_size_choices is None:
    flags.DEFINE_string(**choice_kwargs)
  else:
    flags.DEFINE_enum(enum_values=resnet_size_choices, **choice_kwargs)