resnet_run_loop.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 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

import argparse
import os

import tensorflow as tf  # pylint: disable=g-bad-import-order

from official.resnet import resnet_model
from official.utils.arg_parsers import parsers
from official.utils.logging import hooks_helper


################################################################################
# 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, 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):  # pylint: disable=unused-argument
    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


################################################################################
# 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. It should have one more element
      than `boundary_epochs`, and all elements should have the same type.

  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.
  def exclude_batch_norm(name):
    return 'batch_normalization' not in name
  loss_filter_fn = loss_filter_fn or exclude_batch_norm

  # 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 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.

  Args:
    batch_size: the number of examples processed in each training batch.

  Raises:
    ValueError: if no GPUs are found, or selected batch_size is invalid.
  """
  from tensorflow.python.client import device_lib  # pylint: disable=g-import-not-at-top

  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):
  """Shared main loop for ResNet Models."""

  # 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_between_evals):
    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_between_evals,
                            flags.num_parallel_calls, flags.multi_gpu)

    classifier.train(input_fn=input_fn_train, hooks=train_hooks,
                     max_steps=flags.max_train_steps)

    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)

    # flags.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.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.
    eval_results = classifier.evaluate(input_fn=input_fn_eval,
                                       steps=flags.max_train_steps)
    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=resnet_model.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>' if resnet_size_choices is None else None
    )