resnet.py

# 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 definitions for Residual Networks.

Residual networks ('v1' ResNets) were originally proposed in:
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Deep Residual Learning for Image Recognition. arXiv:1512.03385

The full preactivation 'v2' ResNet variant was introduced by:
[2] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
    Identity Mappings in Deep Residual Networks. arXiv: 1603.05027

The key difference of the full preactivation 'v2' variant compared to the
'v1' variant in [1] is the use of batch normalization before every weight layer
rather than after.
"""

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

import argparse
import os

import tensorflow as tf

from official.utils.arg_parsers import parsers  # pylint: disable=g-bad-import-order
from official.utils.logging import hooks_helper

_BATCH_NORM_DECAY = 0.997
_BATCH_NORM_EPSILON = 1e-5

DEFAULT_VERSION = 2


################################################################################
# Functions for input processing.
################################################################################
def process_record_dataset(dataset, is_training, batch_size, shuffle_buffer,
                           parse_record_fn, num_epochs=1, num_parallel_calls=1,
                           examples_per_epoch=0, multi_gpu=False):
  """Given a Dataset with raw records, parse each record into images and labels,
  and return an iterator over the records.

  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.
    num_parallel_calls: The number of records that are processed in parallel.
      This can be optimized per data set but for generally homogeneous data
      sets, should be approximately the number of available CPU cores.
    examples_per_epoch: The number of examples in the current set that
      are processed each epoch. Note that this is only used for multi-GPU mode,
      and only to handle what will eventually be handled inside of Estimator.
    multi_gpu: Whether this is run multi-GPU. Note that this is only required
      currently to handle the batch leftovers (see below), and can be removed
      when that is handled directly by Estimator.

  Returns:
    Dataset of (image, label) pairs ready for iteration.
  """
  # We prefetch a batch at a time, This can help smooth out the time taken to
  # load input files as we go through shuffling and processing.
  dataset = dataset.prefetch(buffer_size=batch_size)
  if is_training:
    # Shuffle the records. Note that we shuffle before repeating to ensure
    # that the shuffling respects epoch boundaries.
    dataset = dataset.shuffle(buffer_size=shuffle_buffer)

  # If we are training over multiple epochs before evaluating, repeat the
  # dataset for the appropriate number of epochs.
  dataset = dataset.repeat(num_epochs)

  # Currently, if we are using multiple GPUs, we can't pass in uneven batches.
  # (For example, if we have 4 GPUs, the number of examples in each batch
  # must be divisible by 4.) We already ensured this for the batch_size, but
  # we have to additionally ensure that any "leftover" examples-- the remainder
  # examples (total examples % batch_size) that get called a batch for the very
  # last batch of an epoch-- do not raise an error when we try to split them
  # over the GPUs. This will likely be handled by Estimator during replication
  # in the future, but for now, we just drop the leftovers here.
  if multi_gpu:
    total_examples = num_epochs * examples_per_epoch
    dataset = dataset.take(batch_size * (total_examples // batch_size))

  # Parse the raw records into images and labels
  dataset = dataset.map(lambda value: parse_record_fn(value, is_training),
                        num_parallel_calls=num_parallel_calls)

  dataset = dataset.batch(batch_size)

  # 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
  # critical training path.
  dataset = dataset.prefetch(1)

  return dataset


def get_synth_input_fn(height, width, num_channels, num_classes):
  """Returns an input function that returns a dataset with zeroes.

  This is useful in debugging input pipeline performance, as it removes all
  elements of file reading and image preprocessing.

  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

  Returns:
    An input_fn that can be used in place of a real one to return a dataset
    that can be used for iteration.
  """
  def input_fn(is_training, data_dir, batch_size, *args):
    images = tf.zeros((batch_size, height, width, num_channels), tf.float32)
    labels = tf.zeros((batch_size, num_classes), tf.int32)
    return tf.data.Dataset.from_tensors((images, labels)).repeat()

  return input_fn


################################################################################
# Convenience functions for building the ResNet model.
################################################################################
def batch_norm(inputs, training, data_format):
  """Performs a batch normalization using a standard set of parameters."""
  # We set fused=True for a significant performance boost. See
  # https://www.tensorflow.org/performance/performance_guide#common_fused_ops
  return tf.layers.batch_normalization(
      inputs=inputs, axis=1 if data_format == 'channels_first' else 3,
      momentum=_BATCH_NORM_DECAY, epsilon=_BATCH_NORM_EPSILON, center=True,
      scale=True, training=training, fused=True)


def fixed_padding(inputs, kernel_size, data_format):
  """Pads the input along the spatial dimensions independently of input size.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    kernel_size: The kernel to be used in the conv2d or max_pool2d operation.
                 Should be a positive integer.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    A tensor with the same format as the input with the data either intact
    (if kernel_size == 1) or padded (if kernel_size > 1).
  """
  pad_total = kernel_size - 1
  pad_beg = pad_total // 2
  pad_end = pad_total - pad_beg

  if data_format == 'channels_first':
    padded_inputs = tf.pad(inputs, [[0, 0], [0, 0],
                                    [pad_beg, pad_end], [pad_beg, pad_end]])
  else:
    padded_inputs = tf.pad(inputs, [[0, 0], [pad_beg, pad_end],
                                    [pad_beg, pad_end], [0, 0]])
  return padded_inputs


def conv2d_fixed_padding(inputs, filters, kernel_size, strides, data_format):
  """Strided 2-D convolution with explicit padding."""
  # The padding is consistent and is based only on `kernel_size`, not on the
  # dimensions of `inputs` (as opposed to using `tf.layers.conv2d` alone).
  if strides > 1:
    inputs = fixed_padding(inputs, kernel_size, data_format)

  return tf.layers.conv2d(
      inputs=inputs, filters=filters, kernel_size=kernel_size, strides=strides,
      padding=('SAME' if strides == 1 else 'VALID'), use_bias=False,
      kernel_initializer=tf.variance_scaling_initializer(),
      data_format=data_format)


################################################################################
# ResNet block definitions.
################################################################################
def _building_block_v1(inputs, filters, training, projection_shortcut, strides,
    data_format):
  """
  Convolution then batch normalization then ReLU as described by:
    Deep Residual Learning for Image Recognition
    https://arxiv.org/pdf/1512.03385.pdf
    by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Dec 2015.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    filters: The number of filters for the convolutions.
    training: A Boolean for whether the model is in training or inference
      mode. Needed for batch normalization.
    projection_shortcut: The function to use for projection shortcuts
      (typically a 1x1 convolution when downsampling the input).
    strides: The block's stride. If greater than 1, this block will ultimately
      downsample the input.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    The output tensor of the block.
  """
  shortcut = inputs

  if projection_shortcut is not None:
    shortcut = projection_shortcut(inputs)
    shortcut = batch_norm(inputs=shortcut, training=training,
                          data_format=data_format)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=strides,
      data_format=data_format)
  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=1,
      data_format=data_format)
  inputs = batch_norm(inputs, training, data_format)
  inputs += shortcut
  inputs = tf.nn.relu(inputs)

  return inputs


def _building_block_v2(inputs, filters, training, projection_shortcut, strides,
    data_format):
  """
  Batch normalization then ReLu then convolution as described by:
    Identity Mappings in Deep Residual Networks
    https://arxiv.org/pdf/1603.05027.pdf
    by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Jul 2016.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    filters: The number of filters for the convolutions.
    training: A Boolean for whether the model is in training or inference
      mode. Needed for batch normalization.
    projection_shortcut: The function to use for projection shortcuts
      (typically a 1x1 convolution when downsampling the input).
    strides: The block's stride. If greater than 1, this block will ultimately
      downsample the input.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    The output tensor of the block.
  """
  shortcut = inputs
  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)

  # The projection shortcut should come after the first batch norm and ReLU
  # since it performs a 1x1 convolution.
  if projection_shortcut is not None:
    shortcut = projection_shortcut(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=strides,
      data_format=data_format)

  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)
  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=1,
      data_format=data_format)

  return inputs + shortcut


def _bottleneck_block_v1(inputs, filters, training, projection_shortcut,
    strides, data_format):
  """
  Similar to _building_block_v1(), except using the "bottleneck" blocks
  described in:
    Convolution then batch normalization then ReLU as described by:
      Deep Residual Learning for Image Recognition
      https://arxiv.org/pdf/1512.03385.pdf
      by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Dec 2015.
  """
  shortcut = inputs

  if projection_shortcut is not None:
    shortcut = projection_shortcut(inputs)
    shortcut = batch_norm(inputs=shortcut, training=training,
                          data_format=data_format)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=1, strides=1,
      data_format=data_format)
  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=strides,
      data_format=data_format)
  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=4 * filters, kernel_size=1, strides=1,
      data_format=data_format)
  inputs = batch_norm(inputs, training, data_format)
  inputs += shortcut
  inputs = tf.nn.relu(inputs)

  return inputs


def _bottleneck_block_v2(inputs, filters, training, projection_shortcut,
    strides, data_format):
  """
  Similar to _building_block_v2(), except using the "bottleneck" blocks
  described in:
    Convolution then batch normalization then ReLU as described by:
      Deep Residual Learning for Image Recognition
      https://arxiv.org/pdf/1512.03385.pdf
      by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Dec 2015.

  adapted to the ordering conventions of:
    Batch normalization then ReLu then convolution as described by:
      Identity Mappings in Deep Residual Networks
      https://arxiv.org/pdf/1603.05027.pdf
      by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun, Jul 2016.
  """
  shortcut = inputs
  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)

  # The projection shortcut should come after the first batch norm and ReLU
  # since it performs a 1x1 convolution.
  if projection_shortcut is not None:
    shortcut = projection_shortcut(inputs)

  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=1, strides=1,
      data_format=data_format)

  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)
  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=filters, kernel_size=3, strides=strides,
      data_format=data_format)

  inputs = batch_norm(inputs, training, data_format)
  inputs = tf.nn.relu(inputs)
  inputs = conv2d_fixed_padding(
      inputs=inputs, filters=4 * filters, kernel_size=1, strides=1,
      data_format=data_format)

  return inputs + shortcut


def block_layer(inputs, filters, bottleneck, block_fn, blocks, strides,
                training, name, data_format):
  """Creates one layer of blocks for the ResNet model.

  Args:
    inputs: A tensor of size [batch, channels, height_in, width_in] or
      [batch, height_in, width_in, channels] depending on data_format.
    filters: The number of filters for the first convolution of the layer.
    bottleneck: Is the block created a bottleneck block.
    block_fn: The block to use within the model, either `building_block` or
      `bottleneck_block`.
    blocks: The number of blocks contained in the layer.
    strides: The stride to use for the first convolution of the layer. If
      greater than 1, this layer will ultimately downsample the input.
    training: Either True or False, whether we are currently training the
      model. Needed for batch norm.
    name: A string name for the tensor output of the block layer.
    data_format: The input format ('channels_last' or 'channels_first').

  Returns:
    The output tensor of the block layer.
  """

  # Bottleneck blocks end with 4x the number of filters as they start with
  filters_out = filters * 4 if bottleneck else filters

  def projection_shortcut(inputs):
    return conv2d_fixed_padding(
        inputs=inputs, filters=filters_out, kernel_size=1, strides=strides,
        data_format=data_format)

  # Only the first block per block_layer uses projection_shortcut and strides
  inputs = block_fn(inputs, filters, training, projection_shortcut, strides,
                    data_format)

  for _ in range(1, blocks):
    inputs = block_fn(inputs, filters, training, None, 1, data_format)

  return tf.identity(inputs, name)


class Model(object):
  """Base class for building the Resnet Model.
  """

  def __init__(self, resnet_size, bottleneck, num_classes, num_filters,
               kernel_size,
               conv_stride, first_pool_size, first_pool_stride,
               second_pool_size, second_pool_stride, block_sizes, block_strides,
               final_size, version=DEFAULT_VERSION, data_format=None):
    """Creates a model for classifying an image.

    Args:
      resnet_size: A single integer for the size of the ResNet model.
      bottleneck: Use regular blocks or bottleneck blocks.
      num_classes: The number of classes used as labels.
      num_filters: The number of filters to use for the first block layer
        of the model. This number is then doubled for each subsequent block
        layer.
      kernel_size: The kernel size to use for convolution.
      conv_stride: stride size for the initial convolutional layer
      first_pool_size: Pool size to be used for the first pooling layer.
        If none, the first pooling layer is skipped.
      first_pool_stride: stride size for the first pooling layer. Not used
        if first_pool_size is None.
      second_pool_size: Pool size to be used for the second pooling layer.
      second_pool_stride: stride size for the final pooling layer
      block_sizes: A list containing n values, where n is the number of sets of
        block layers desired. Each value should be the number of blocks in the
        i-th set.
      block_strides: List of integers representing the desired stride size for
        each of the sets of block layers. Should be same length as block_sizes.
      final_size: The expected size of the model after the second pooling.
      version: Integer representing which version of the ResNet network to use.
        See README for details. Valid values: [1, 2]
      data_format: Input format ('channels_last', 'channels_first', or None).
        If set to None, the format is dependent on whether a GPU is available.
    """
    self.resnet_size = resnet_size

    if not data_format:
      data_format = (
          'channels_first' if tf.test.is_built_with_cuda() else 'channels_last')

    self.resnet_version = version
    if version not in (1, 2):
      raise ValueError(
          "Resnet version should be 1 or 2. See README for citations.")

    self.bottleneck = bottleneck
    if bottleneck:
      if version == 1:
        self.block_fn = _bottleneck_block_v1
      else:
        self.block_fn = _bottleneck_block_v2
    else:
      if version == 1:
        self.block_fn = _building_block_v1
      else:
        self.block_fn = _building_block_v2

    self.data_format = data_format
    self.num_classes = num_classes
    self.num_filters = num_filters
    self.kernel_size = kernel_size
    self.conv_stride = conv_stride
    self.first_pool_size = first_pool_size
    self.first_pool_stride = first_pool_stride
    self.second_pool_size = second_pool_size
    self.second_pool_stride = second_pool_stride
    self.block_sizes = block_sizes
    self.block_strides = block_strides
    self.final_size = final_size

  def __call__(self, inputs, training):
    """Add operations to classify a batch of input images.

    Args:
      inputs: A Tensor representing a batch of input images.
      training: A boolean. Set to True to add operations required only when
        training the classifier.

    Returns:
      A logits Tensor with shape [<batch_size>, self.num_classes].
    """

    if self.data_format == 'channels_first':
      # Convert the inputs from channels_last (NHWC) to channels_first (NCHW).
      # This provides a large performance boost on GPU. See
      # https://www.tensorflow.org/performance/performance_guide#data_formats
      inputs = tf.transpose(inputs, [0, 3, 1, 2])

    inputs = conv2d_fixed_padding(
        inputs=inputs, filters=self.num_filters, kernel_size=self.kernel_size,
        strides=self.conv_stride, data_format=self.data_format)
    inputs = tf.identity(inputs, 'initial_conv')

    if self.first_pool_size:
      inputs = tf.layers.max_pooling2d(
          inputs=inputs, pool_size=self.first_pool_size,
          strides=self.first_pool_stride, padding='SAME',
          data_format=self.data_format)
      inputs = tf.identity(inputs, 'initial_max_pool')

    for i, num_blocks in enumerate(self.block_sizes):
      num_filters = self.num_filters * (2**i)
      inputs = block_layer(
          inputs=inputs, filters=num_filters, bottleneck=self.bottleneck,
          block_fn=self.block_fn, blocks=num_blocks,
          strides=self.block_strides[i], training=training,
          name='block_layer{}'.format(i + 1), data_format=self.data_format)

    inputs = batch_norm(inputs, training, self.data_format)
    inputs = tf.nn.relu(inputs)
    inputs = tf.layers.average_pooling2d(
        inputs=inputs, pool_size=self.second_pool_size,
        strides=self.second_pool_stride, padding='VALID',
        data_format=self.data_format)
    inputs = tf.identity(inputs, 'final_avg_pool')

    inputs = tf.reshape(inputs, [-1, self.final_size])
    inputs = tf.layers.dense(inputs=inputs, units=self.num_classes)
    inputs = tf.identity(inputs, 'final_dense')
    return inputs


################################################################################
# Functions for running training/eval/validation loops for the model.
################################################################################
def learning_rate_with_decay(
    batch_size, batch_denom, num_images, boundary_epochs, decay_rates):
  """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
      for scaling the learning rate. Should be the same length as
      boundary_epochs.

  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.
  """
  initial_learning_rate = 0.1 * batch_size / batch_denom
  batches_per_epoch = num_images / batch_size

  # Multiply the learning rate by 0.1 at 100, 150, and 200 epochs.
  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):
    global_step = tf.cast(global_step, tf.int32)
    return tf.train.piecewise_constant(global_step, boundaries, vals)

  return learning_rate_fn


def resnet_model_fn(features, labels, mode, model_class,
                    resnet_size, weight_decay, learning_rate_fn, momentum,
                    data_format, version, loss_filter_fn=None, multi_gpu=False):
  """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.
    version: Integer representing which version of the ResNet network to use.
      See README for details. Valid values: [1, 2]
    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.
    multi_gpu: If True, wrap the optimizer in a TowerOptimizer suitable for
      data-parallel distribution across multiple GPUs.

  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)

  model = model_class(resnet_size, data_format, version=version)
  logits = model(features, mode == tf.estimator.ModeKeys.TRAIN)

  predictions = {
      'classes': tf.argmax(logits, axis=1),
      'probabilities': tf.nn.softmax(logits, name='softmax_tensor')
  }

  if mode == tf.estimator.ModeKeys.PREDICT:
    return tf.estimator.EstimatorSpec(mode=mode, predictions=predictions)

  # Calculate loss, which includes softmax cross entropy and L2 regularization.
  cross_entropy = tf.losses.softmax_cross_entropy(
      logits=logits, onehot_labels=labels)

  # 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.
  if not loss_filter_fn:
    def loss_filter_fn(name):
      return 'batch_normalization' not in name

  # Add weight decay to the loss.
  loss = cross_entropy + weight_decay * tf.add_n(
      [tf.nn.l2_loss(v) for v in tf.trainable_variables()
       if loss_filter_fn(v.name)])

  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,
        momentum=momentum)

    # If we are running multi-GPU, we need to wrap the optimizer.
    if multi_gpu:
      optimizer = tf.contrib.estimator.TowerOptimizer(optimizer)

    update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
    train_op = tf.group(optimizer.minimize(loss, global_step), update_ops)
  else:
    train_op = None

  accuracy = tf.metrics.accuracy(
      tf.argmax(labels, axis=1), predictions['classes'])
  metrics = {'accuracy': accuracy}

  # Create a tensor named train_accuracy for logging purposes
  tf.identity(accuracy[1], name='train_accuracy')
  tf.summary.scalar('train_accuracy', accuracy[1])

  return tf.estimator.EstimatorSpec(
      mode=mode,
      predictions=predictions,
      loss=loss,
      train_op=train_op,
      eval_metric_ops=metrics)


def validate_batch_size_for_multi_gpu(batch_size):
  """For multi-gpu, batch-size must be a multiple of the number of
  available GPUs.

  Note that this should eventually be handled by replicate_model_fn
  directly. Multi-GPU support is currently experimental, however,
  so doing the work here until that feature is in place.
  """
  from tensorflow.python.client import device_lib

  local_device_protos = device_lib.list_local_devices()
  num_gpus = sum([1 for d in local_device_protos if d.device_type == 'GPU'])
  if not num_gpus:
    raise ValueError('Multi-GPU mode was specified, but no GPUs '
      'were found. To use CPU, run without --multi_gpu.')

  remainder = batch_size % num_gpus
  if remainder:
    err = ('When running with multiple GPUs, batch size '
      'must be a multiple of the number of available GPUs. '
      'Found {} GPUs with a batch size of {}; try --batch_size={} instead.'
      ).format(num_gpus, batch_size, batch_size - remainder)
    raise ValueError(err)


def resnet_main(flags, model_function, input_function):
  # Using the Winograd non-fused algorithms provides a small performance boost.
  os.environ['TF_ENABLE_WINOGRAD_NONFUSED'] = '1'

  if flags.multi_gpu:
    validate_batch_size_for_multi_gpu(flags.batch_size)

    # There are two steps required if using multi-GPU: (1) wrap the model_fn,
    # and (2) wrap the optimizer. The first happens here, and (2) happens
    # in the model_fn itself when the optimizer is defined.
    model_function = tf.contrib.estimator.replicate_model_fn(
        model_function,
        loss_reduction=tf.losses.Reduction.MEAN)

  # 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.inter_op_parallelism_threads,
      intra_op_parallelism_threads=flags.intra_op_parallelism_threads,
      allow_soft_placement=True)

  # Set up a RunConfig to save checkpoint and set session config.
  run_config = tf.estimator.RunConfig().replace(save_checkpoints_secs=1e9,
                                                session_config=session_config)
  classifier = tf.estimator.Estimator(
      model_fn=model_function, model_dir=flags.model_dir, config=run_config,
      params={
          'resnet_size': flags.resnet_size,
          'data_format': flags.data_format,
          'batch_size': flags.batch_size,
          'multi_gpu': flags.multi_gpu,
          'version': flags.version,
      })

  for _ in range(flags.train_epochs // flags.epochs_per_eval):
    train_hooks = hooks_helper.get_train_hooks(flags.hooks, batch_size=flags.batch_size)

    print('Starting a training cycle.')

    def input_fn_train():
      return input_function(True, flags.data_dir, flags.batch_size,
                            flags.epochs_per_eval, flags.num_parallel_calls,
                            flags.multi_gpu)

    classifier.train(input_fn=input_fn_train, hooks=train_hooks)

    print('Starting to evaluate.')
    # Evaluate the model and print results
    def input_fn_eval():
      return input_function(False, flags.data_dir, flags.batch_size,
                            1, flags.num_parallel_calls, flags.multi_gpu)

    eval_results = classifier.evaluate(input_fn=input_fn_eval)
    print(eval_results)


class ResnetArgParser(argparse.ArgumentParser):
  """Arguments for configuring and running a Resnet Model.
  """

  def __init__(self, resnet_size_choices=None):
    super(ResnetArgParser, self).__init__(parents=[
        parsers.BaseParser(),
        parsers.PerformanceParser(),
        parsers.ImageModelParser(),
    ])

    self.add_argument(
        '--version', '-v', type=int, choices=[1, 2],
        default=DEFAULT_VERSION,
        help="Version of ResNet. (1 or 2) See README.md for details."
    )

    self.add_argument(
        '--resnet_size', '-rs', type=int, default=50,
        choices=resnet_size_choices,
        help='[default: %(default)s]The size of the ResNet model to use.',
        metavar='<RS>'
    )