path: root/tensorflow/contrib/learn/python/learn/graph_actions.py

# Copyright 2016 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.
# ==============================================================================

"""High level operations on graphs."""

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

import itertools
import sys
import threading
import time

import numpy as np

from six import reraise

from tensorflow.contrib.framework import load_variable
from tensorflow.contrib.framework.python.ops import ops as contrib_ops
from tensorflow.contrib.framework.python.ops import variables as contrib_variables
from tensorflow.contrib.learn.python.learn import monitors as monitors_lib
from tensorflow.core.framework import summary_pb2
from tensorflow.core.protobuf import saver_pb2
from tensorflow.python.client import session as tf_session
from tensorflow.python.framework import errors
from tensorflow.python.framework import ops
from tensorflow.python.ops import control_flow_ops
from tensorflow.python.ops import data_flow_ops
from tensorflow.python.ops import logging_ops
from tensorflow.python.ops import resources
from tensorflow.python.ops import variables
from tensorflow.python.platform import tf_logging as logging
from tensorflow.python.training import basic_session_run_hooks
from tensorflow.python.training import coordinator
from tensorflow.python.training import monitored_session
from tensorflow.python.training import queue_runner
from tensorflow.python.training import saver as tf_saver
from tensorflow.python.training import session_manager as session_manager_lib
from tensorflow.python.training import summary_io
from tensorflow.python.training import supervisor as tf_supervisor

# Singleton for SummaryWriter per logdir folder.
_SUMMARY_WRITERS = {}

# Lock protecting _SUMMARY_WRITERS
_summary_writer_lock = threading.Lock()


def clear_summary_writers():
  """Clear cached summary writers. Currently only used for unit tests."""
  return summary_io.SummaryWriterCache.clear()


def get_summary_writer(logdir):
  """Returns single SummaryWriter per logdir in current run.

  Args:
    logdir: str, folder to write summaries.

  Returns:
    Existing `SummaryWriter` object or new one if never wrote to given
    directory.
  """
  return summary_io.SummaryWriterCache.get(logdir)


def _make_saver(graph, keep_checkpoint_max=5):
  vars_to_save = (graph.get_collection(ops.GraphKeys.GLOBAL_VARIABLES) +
                  graph.get_collection(ops.GraphKeys.SAVEABLE_OBJECTS))
  if vars_to_save:
    return tf_saver.Saver(vars_to_save,
                          sharded=True,
                          max_to_keep=keep_checkpoint_max)
  else:
    return None


def _restore_from_checkpoint(session, graph, checkpoint_path, saver=None):
  logging.info('Loading model from checkpoint: %s.', checkpoint_path)
  saver = saver or _make_saver(graph)
  if saver:
    saver.restore(session, checkpoint_path)
  else:
    logging.info('No variables found in graph, not creating Saver() object.')


def _run_with_monitors(session, step, tensors, feed_dict, monitors):
  """Runs session for given tensors with monitor callbacks."""
  for monitor in monitors:
    tensors += monitor.step_begin(step)
  tensors = list(set(tensors))

  outputs = session.run(tensors, feed_dict=feed_dict)
  outputs = dict(zip(
      [t.name if isinstance(t, ops.Tensor) else t for t in tensors],
      outputs))

  should_stop = False
  for monitor in monitors:
    induce_stop = monitor.step_end(step, outputs)
    should_stop = should_stop or induce_stop
  return outputs, should_stop


def _monitored_train(graph,
                     output_dir,
                     train_op,
                     loss_op,
                     global_step_tensor=None,
                     init_op=None,
                     init_feed_dict=None,
                     init_fn=None,
                     log_every_steps=10,
                     supervisor_is_chief=True,
                     supervisor_master='',
                     supervisor_save_model_secs=600,
                     supervisor_save_model_steps=None,
                     keep_checkpoint_max=5,
                     supervisor_save_summaries_secs=None,
                     supervisor_save_summaries_steps=100,
                     feed_fn=None,
                     steps=None,
                     fail_on_nan_loss=True,
                     hooks=None,
                     max_steps=None):
  """Train a model via monitored_session.

  Given `graph`, a directory to write outputs to (`output_dir`), and some ops,
  run a training loop. The given `train_op` performs one step of training on the
  model. The `loss_op` represents the objective function of the training. It is
  expected to increment the `global_step_tensor`, a scalar integer tensor
  counting training steps. This function uses `Supervisor` to initialize the
  graph (from a checkpoint if one is available in `output_dir`), write summaries
  defined in the graph, and write regular checkpoints as defined by
  `supervisor_save_model_secs`.

  Training continues until `global_step_tensor` evaluates to `max_steps`, or, if
  `fail_on_nan_loss`, until `loss_op` evaluates to `NaN`. In that case the
  program is terminated with exit code 1.

  Args:
    graph: A graph to train. It is expected that this graph is not in use
      elsewhere.
    output_dir: A directory to write outputs to.
    train_op: An op that performs one training step when run.
    loss_op: A scalar loss tensor.
    global_step_tensor: A tensor representing the global step. If none is given,
      one is extracted from the graph using the same logic as in `Supervisor`.
    init_op: An op that initializes the graph. If `None`, use `Supervisor`'s
      default.
    init_feed_dict: A dictionary that maps `Tensor` objects to feed values.
      This feed dictionary will be used when `init_op` is evaluated.
    init_fn: Optional callable passed to Supervisor to initialize the model.
    log_every_steps: Output logs regularly. The logs contain timing data and the
      current loss. A `0` or negative value disables logging.
    supervisor_is_chief: Whether the current process is the chief supervisor in
      charge of restoring the model and running standard services.
    supervisor_master: The master string to use when preparing the session.
    supervisor_save_model_secs: Save checkpoints every this many seconds. Can
        not be specified with `supervisor_save_model_steps`.
    supervisor_save_model_steps: Save checkpoints every this many steps. Can not
        be specified with `supervisor_save_model_secs`.
    keep_checkpoint_max: The maximum number of recent checkpoint files to
      keep. As new files are created, older files are deleted. If None or 0,
      all checkpoint files are kept. This is simply passed as the max_to_keep
      arg to `tf.Saver` constructor.
    supervisor_save_summaries_secs: Save summaries every
      `supervisor_save_summaries_secs` seconds when training.
    supervisor_save_summaries_steps: Save summaries every
      `supervisor_save_summaries_steps` steps when training. Exactly one of
      `supervisor_save_model_steps` and `supervisor_save_model_secs` should be
      specified, and the other should be None.
    feed_fn: A function that is called every iteration to produce a `feed_dict`
      passed to `session.run` calls. Optional.
    steps: Trains for this many steps (e.g. current global step + `steps`).
    fail_on_nan_loss: If true, raise `NanLossDuringTrainingError` if `loss_op`
      evaluates to `NaN`. If false, continue training as if nothing happened.
    hooks: List of `SessionRunHook` subclass instances. Used for callbacks
      inside the training loop.
    max_steps: Number of total steps for which to train model. If `None`,
      train forever. Two calls fit(steps=100) means 200 training iterations.
      On the other hand two calls of fit(max_steps=100) means, second call
      will not do any iteration since first call did all 100 steps.

  Returns:
    The final loss value.

  Raises:
    ValueError: If `output_dir`, `train_op`, `loss_op`, or `global_step_tensor`
      is not provided. See `tf.contrib.framework.get_global_step` for how we
      look up the latter if not provided explicitly.
    NanLossDuringTrainingError: If `fail_on_nan_loss` is `True`, and loss ever
      evaluates to `NaN`.
    ValueError: If both `steps` and `max_steps` are not `None`.
  """
  if (steps is not None) and (max_steps is not None):
    raise ValueError('Can not provide both steps and max_steps.')
  if not output_dir:
    raise ValueError('Output directory should be non-empty %s.' % output_dir)
  if train_op is None:
    raise ValueError('Missing train_op.')
  if loss_op is None:
    raise ValueError('Missing loss_op.')
  if hooks is None:
    hooks = []
  if not isinstance(hooks, list):
    raise ValueError('Hooks should be a list.')
  with graph.as_default():
    global_step_tensor = contrib_variables.assert_or_get_global_step(
        graph, global_step_tensor)
  if global_step_tensor is None:
    raise ValueError('No "global_step" was provided or found in the graph.')

  if max_steps is not None:
    try:
      start_step = load_variable(output_dir, global_step_tensor.name)
      if max_steps <= start_step:
        logging.info('Skipping training since max_steps has already saved.')
        return None
    except:  # pylint: disable=bare-except
      pass

  # Adapted SessionRunHooks such as ExportMonitor depend on the
  # CheckpointSaverHook to be executed before they should be executed.
  # The `hooks` param comprises of deprecated monitor hooks
  # (such as ExportMonitor). Appending them after the basic_session_run_hooks.
  all_hooks = []
  with graph.as_default():
    all_hooks.append(basic_session_run_hooks.NanTensorHook(
        loss_op, fail_on_nan_loss=fail_on_nan_loss))
    if log_every_steps > 0:
      all_hooks.append(basic_session_run_hooks.LoggingTensorHook({
          'loss': loss_op.name,
          'step': global_step_tensor.name
      }, every_n_iter=log_every_steps))

    def make_saver():
      return tf_saver.Saver(
          sharded=True, max_to_keep=keep_checkpoint_max, defer_build=True,
          write_version=saver_pb2.SaverDef.V1)

    scaffold = monitored_session.Scaffold(
        init_op=init_op,
        init_feed_dict=init_feed_dict,
        init_fn=init_fn,
        saver=monitored_session.Scaffold.get_or_default('saver',
                                                        ops.GraphKeys.SAVERS,
                                                        make_saver))

    if not supervisor_is_chief:
      session_creator = monitored_session.WorkerSessionCreator(
          scaffold=scaffold,
          master=supervisor_master)
    else:
      session_creator = monitored_session.ChiefSessionCreator(
          scaffold=scaffold,
          checkpoint_dir=output_dir,
          master=supervisor_master)
      summary_writer = summary_io.SummaryWriterCache.get(output_dir)
      all_hooks.append(
          basic_session_run_hooks.StepCounterHook(
              summary_writer=summary_writer))
      all_hooks.append(
          basic_session_run_hooks.SummarySaverHook(
              save_secs=supervisor_save_summaries_secs,
              save_steps=supervisor_save_summaries_steps,
              summary_writer=summary_writer,
              scaffold=scaffold))
      if (supervisor_save_model_secs is not None
          or supervisor_save_model_steps is not None):
        all_hooks.append(
            basic_session_run_hooks.CheckpointSaverHook(
                output_dir,
                save_secs=supervisor_save_model_secs,
                save_steps=supervisor_save_model_steps,
                scaffold=scaffold))

    if steps is not None or max_steps is not None:
      all_hooks.append(basic_session_run_hooks.StopAtStepHook(steps, max_steps))
    all_hooks.extend(hooks)

    with monitored_session.MonitoredSession(
        session_creator=session_creator,
        hooks=all_hooks) as super_sess:
      loss = None
      while not super_sess.should_stop():
        _, loss = super_sess.run([train_op, loss_op], feed_fn() if feed_fn else
                                 None)
    summary_io.SummaryWriterCache.clear()
    return loss


# TODO(ispir): Deprecate train in favor of supervised_train
def train(graph,
          output_dir,
          train_op,
          loss_op,
          global_step_tensor=None,
          init_op=None,
          init_feed_dict=None,
          init_fn=None,
          log_every_steps=10,
          supervisor_is_chief=True,
          supervisor_master='',
          supervisor_save_model_secs=600,
          keep_checkpoint_max=5,
          supervisor_save_summaries_steps=100,
          feed_fn=None,
          steps=None,
          fail_on_nan_loss=True,
          monitors=None,
          max_steps=None):
  """Train a model.

  Given `graph`, a directory to write outputs to (`output_dir`), and some ops,
  run a training loop. The given `train_op` performs one step of training on the
  model. The `loss_op` represents the objective function of the training. It is
  expected to increment the `global_step_tensor`, a scalar integer tensor
  counting training steps. This function uses `Supervisor` to initialize the
  graph (from a checkpoint if one is available in `output_dir`), write summaries
  defined in the graph, and write regular checkpoints as defined by
  `supervisor_save_model_secs`.

  Training continues until `global_step_tensor` evaluates to `max_steps`, or, if
  `fail_on_nan_loss`, until `loss_op` evaluates to `NaN`. In that case the
  program is terminated with exit code 1.

  Args:
    graph: A graph to train. It is expected that this graph is not in use
      elsewhere.
    output_dir: A directory to write outputs to.
    train_op: An op that performs one training step when run.
    loss_op: A scalar loss tensor.
    global_step_tensor: A tensor representing the global step. If none is given,
      one is extracted from the graph using the same logic as in `Supervisor`.
    init_op: An op that initializes the graph. If `None`, use `Supervisor`'s
      default.
    init_feed_dict: A dictionary that maps `Tensor` objects to feed values.
      This feed dictionary will be used when `init_op` is evaluated.
    init_fn: Optional callable passed to Supervisor to initialize the model.
    log_every_steps: Output logs regularly. The logs contain timing data and the
      current loss.
    supervisor_is_chief: Whether the current process is the chief supervisor in
      charge of restoring the model and running standard services.
    supervisor_master: The master string to use when preparing the session.
    supervisor_save_model_secs: Save a checkpoint every
      `supervisor_save_model_secs` seconds when training.
    keep_checkpoint_max: The maximum number of recent checkpoint files to
      keep. As new files are created, older files are deleted. If None or 0,
      all checkpoint files are kept. This is simply passed as the max_to_keep
      arg to tf.Saver constructor.
    supervisor_save_summaries_steps: Save summaries every
      `supervisor_save_summaries_steps` seconds when training.
    feed_fn: A function that is called every iteration to produce a `feed_dict`
      passed to `session.run` calls. Optional.
    steps: Trains for this many steps (e.g. current global step + `steps`).
    fail_on_nan_loss: If true, raise `NanLossDuringTrainingError` if `loss_op`
      evaluates to `NaN`. If false, continue training as if nothing happened.
    monitors: List of `BaseMonitor` subclass instances. Used for callbacks
      inside the training loop.
    max_steps: Number of total steps for which to train model. If `None`,
      train forever. Two calls fit(steps=100) means 200 training iterations.
      On the other hand two calls of fit(max_steps=100) means, second call
      will not do any iteration since first call did all 100 steps.

  Returns:
    The final loss value.

  Raises:
    ValueError: If `output_dir`, `train_op`, `loss_op`, or `global_step_tensor`
      is not provided. See `tf.contrib.framework.get_global_step` for how we
      look up the latter if not provided explicitly.
    NanLossDuringTrainingError: If `fail_on_nan_loss` is `True`, and loss ever
      evaluates to `NaN`.
    ValueError: If both `steps` and `max_steps` are not `None`.
  """
  while True:
    try:
      return _train_internal(graph,
                             output_dir,
                             train_op,
                             loss_op,
                             global_step_tensor,
                             init_op,
                             init_feed_dict,
                             init_fn,
                             log_every_steps,
                             supervisor_is_chief,
                             supervisor_master,
                             supervisor_save_model_secs,
                             keep_checkpoint_max,
                             supervisor_save_summaries_steps,
                             feed_fn,
                             steps,
                             fail_on_nan_loss,
                             monitors,
                             max_steps)
    except errors.AbortedError:
      # Happens when PS restarts, keep training.
      logging.warning('Training got Aborted error. Keep training.')


def _train_internal(graph,
                    output_dir,
                    train_op,
                    loss_op,
                    global_step_tensor,
                    init_op,
                    init_feed_dict,
                    init_fn,
                    log_every_steps,
                    supervisor_is_chief,
                    supervisor_master,
                    supervisor_save_model_secs,
                    keep_checkpoint_max,
                    supervisor_save_summaries_steps,
                    feed_fn,
                    steps,
                    fail_on_nan_loss,
                    monitors,
                    max_steps):
  """See train."""
  if (steps is not None) and (max_steps is not None):
    raise ValueError('Can not provide both steps and max_steps.')
  if not output_dir:
    raise ValueError('Output directory should be non-empty %s.' % output_dir)
  if train_op is None:
    raise ValueError('Missing train_op.')
  if loss_op is None:
    raise ValueError('Missing loss_op.')

  with graph.as_default():
    global_step_tensor = contrib_variables.assert_or_get_global_step(
        graph, global_step_tensor)
    if global_step_tensor is None:
      raise ValueError('No "global_step" was provided or found in the graph.')

    # Get current step.
    try:
      start_step = load_variable(output_dir, global_step_tensor.name)
    except (errors.NotFoundError, ValueError):
      start_step = 0

    summary_writer = (get_summary_writer(output_dir)
                      if supervisor_is_chief else None)

    # Add default chief monitors if none were provided.
    if not monitors:
      monitors = monitors_lib.get_default_monitors(
          loss_op=loss_op,
          summary_op=logging_ops.get_summary_op(),
          save_summary_steps=supervisor_save_summaries_steps,
          summary_writer=summary_writer) if supervisor_is_chief else []

    # TODO(ipolosukhin): Replace all functionality of Supervisor
    # with Chief-Exclusive Monitors.
    if not supervisor_is_chief:
      # Prune list of monitor to the ones runnable on all workers.
      monitors = [monitor for monitor in monitors if monitor.run_on_all_workers]

    if max_steps is None:
      max_steps = (start_step + steps) if steps else None
    # Start monitors, can create graph parts.
    for monitor in monitors:
      monitor.begin(max_steps=max_steps)

  supervisor = tf_supervisor.Supervisor(
      graph,
      init_op=init_op or tf_supervisor.Supervisor.USE_DEFAULT,
      init_feed_dict=init_feed_dict,
      is_chief=supervisor_is_chief,
      logdir=output_dir,
      saver=_make_saver(graph, keep_checkpoint_max),
      global_step=global_step_tensor,
      summary_op=None,
      summary_writer=summary_writer,
      save_model_secs=supervisor_save_model_secs,
      init_fn=init_fn)
  session = supervisor.PrepareSession(master=supervisor_master,
                                      start_standard_services=True)
  supervisor.StartQueueRunners(session)

  with session:
    get_current_step = lambda: session.run(global_step_tensor)

    start_step = get_current_step()
    last_step = start_step
    last_log_step = start_step
    loss_value = None
    logging.info('Training steps [%d,%s)', last_step, 'inf'
                 if max_steps is None else str(max_steps))

    excinfo = None
    try:
      while not supervisor.ShouldStop() and (
          (max_steps is None) or (last_step < max_steps)):
        start_time = time.time()
        feed_dict = feed_fn() if feed_fn is not None else None

        outputs, should_stop = _run_with_monitors(
            session, last_step + 1, [train_op, loss_op], feed_dict, monitors)

        loss_value = outputs[loss_op.name]
        if np.isnan(loss_value):
          failure_message = 'Model diverged with loss = NaN.'
          if fail_on_nan_loss:
            logging.error(failure_message)
            raise monitors_lib.NanLossDuringTrainingError()
          else:
            logging.warning(failure_message)

        if should_stop:
          break

        this_step = get_current_step()

        if this_step <= last_step:
          logging.error(
              'Global step was not incremented by train op at step %s'
              ': new step %d', last_step, this_step)

        last_step = this_step
        is_last_step = (max_steps is not None) and (last_step >= max_steps)
        if is_last_step or (last_step - last_log_step >= log_every_steps):
          logging.info(
              'training step %d, loss = %.5f (%.3f sec/batch).',
              last_step, loss_value, float(time.time() - start_time))
          last_log_step = last_step
    except errors.OutOfRangeError as e:
      logging.warn('Got exception during tf.learn training loop possibly '
                   'due to exhausted input queue %s.', e)
    except StopIteration:
      logging.info('Exhausted input iterarator.')
    except BaseException as e:  # pylint: disable=broad-except
      # Hold on to any other exceptions while we try recording a final
      # checkpoint and summary.
      excinfo = sys.exc_info()
    finally:
      try:
        # Call supervisor.Stop() from within a try block because it re-raises
        # exceptions thrown by the supervised threads.
        supervisor.Stop(close_summary_writer=False)

        # Save one last checkpoint and summaries
        # TODO(wicke): This should be handled by Supervisor

        # In case we encountered an exception in the try block before we updated
        # last_step, update it here (again).
        last_step = get_current_step()
        if supervisor_is_chief:
          ckpt_path = supervisor.save_path
          logging.info('Saving checkpoint for step %d to checkpoint: %s.',
                       last_step, ckpt_path)
          supervisor.saver.save(session, ckpt_path, global_step=last_step)

          # Finish monitors.
          for monitor in monitors:
            monitor.end()

      # catch OutOfRangeError which is thrown when queue is out of data (and for
      # other reasons as well).
      except errors.OutOfRangeError as e:
        logging.warn('OutOfRangeError in tf.learn final checkpoint possibly '
                     'due to exhausted input queue. Note: summary_op is not '
                     'expected to trigger dequeues. %s.', e)
      except BaseException as e:  # pylint: disable=broad-except
        # If we don't already have an exception to re-raise, raise this one.
        if not excinfo:
          raise
        # Otherwise, log this one and raise the other in the finally block.
        logging.error('Got exception during tf.learn final checkpoint %s.', e)
      finally:
        if excinfo:
          reraise(*excinfo)
    return loss_value


def _get_first_op_from_collection(collection_name):
  elements = ops.get_collection(collection_name)
  if elements:
    return elements[0]
  return None


def _get_saver():
  """Lazy init and return saver."""
  saver = _get_first_op_from_collection(ops.GraphKeys.SAVERS)
  if saver is None and variables.global_variables():
    saver = tf_saver.Saver()
    ops.add_to_collection(ops.GraphKeys.SAVERS, saver)
  return saver


def _get_ready_op():
  ready_op = _get_first_op_from_collection(ops.GraphKeys.READY_OP)
  if ready_op is None:
    ready_op = variables.report_uninitialized_variables()
    ops.add_to_collection(ops.GraphKeys.READY_OP, ready_op)
  return ready_op


def _get_local_init_op():
  local_init_op = _get_first_op_from_collection(
      ops.GraphKeys.LOCAL_INIT_OP)
  if local_init_op is None:
    op_list = [variables.local_variables_initializer(),
               data_flow_ops.initialize_all_tables()]
    if op_list:
      local_init_op = control_flow_ops.group(*op_list)
      ops.add_to_collection(ops.GraphKeys.LOCAL_INIT_OP, local_init_op)
  return local_init_op


def _eval_results_to_str(eval_results):
  return ', '.join('%s = %s' % (k, v) for k, v in sorted(eval_results.items()))


def _write_summary_results(output_dir, eval_results, current_global_step):
  """Writes eval results into summary file in given dir."""
  logging.info('Saving evaluation summary for step %d: %s', current_global_step,
               _eval_results_to_str(eval_results))
  summary_writer = get_summary_writer(output_dir)
  summary = summary_pb2.Summary()
  for key in eval_results:
    if eval_results[key] is None:
      continue
    value = summary.value.add()
    value.tag = key
    if (isinstance(eval_results[key], np.float32) or
        isinstance(eval_results[key], float)):
      value.simple_value = float(eval_results[key])
    else:
      logging.warn('Skipping summary for %s, must be a float or np.float32.',
                   key)
  summary_writer.add_summary(summary, current_global_step)
  summary_writer.flush()


def evaluate(graph,
             output_dir,
             checkpoint_path,
             eval_dict,
             update_op=None,
             global_step_tensor=None,
             supervisor_master='',
             log_every_steps=10,
             feed_fn=None,
             max_steps=None):
  """Evaluate a model loaded from a checkpoint.

  Given `graph`, a directory to write summaries to (`output_dir`), a checkpoint
  to restore variables from, and a `dict` of `Tensor`s to evaluate, run an eval
  loop for `max_steps` steps, or until an exception (generally, an
  end-of-input signal from a reader operation) is raised from running
  `eval_dict`.

  In each step of evaluation, all tensors in the `eval_dict` are evaluated, and
  every `log_every_steps` steps, they are logged. At the very end of evaluation,
  a summary is evaluated (finding the summary ops using `Supervisor`'s logic)
  and written to `output_dir`.

  Args:
    graph: A `Graph` to train. It is expected that this graph is not in use
      elsewhere.
    output_dir: A string containing the directory to write a summary to.
    checkpoint_path: A string containing the path to a checkpoint to restore.
      Can be `None` if the graph doesn't require loading any variables.
    eval_dict: A `dict` mapping string names to tensors to evaluate. It is
      evaluated in every logging step. The result of the final evaluation is
      returned. If `update_op` is None, then it's evaluated in every step. If
      `max_steps` is `None`, this should depend on a reader that will raise an
      end-of-input exception when the inputs are exhausted.
    update_op: A `Tensor` which is run in every step.
    global_step_tensor: A `Variable` containing the global step. If `None`,
      one is extracted from the graph using the same logic as in `Supervisor`.
      Used to place eval summaries on training curves.
    supervisor_master: The master string to use when preparing the session.
    log_every_steps: Integer. Output logs every `log_every_steps` evaluation
      steps. The logs contain the `eval_dict` and timing information.
    feed_fn: A function that is called every iteration to produce a `feed_dict`
      passed to `session.run` calls. Optional.
    max_steps: Integer. Evaluate `eval_dict` this many times.

  Returns:
    A tuple `(eval_results, global_step)`:
    eval_results: A `dict` mapping `string` to numeric values (`int`, `float`)
      that are the result of running eval_dict in the last step. `None` if no
      eval steps were run.
    global_step: The global step this evaluation corresponds to.

  Raises:
    ValueError: if `output_dir` is empty.
  """
  if not output_dir:
    raise ValueError('Output directory should be non-empty %s.' % output_dir)
  with graph.as_default():
    global_step_tensor = contrib_variables.assert_or_get_global_step(
        graph, global_step_tensor)

    # Create or get summary op, global_step and saver.
    saver = _get_saver()
    local_init_op = _get_local_init_op()
    ready_op = _get_ready_op()

    session_manager = session_manager_lib.SessionManager(
        local_init_op=local_init_op,
        ready_op=ready_op)
    session, initialized = session_manager.recover_session(
        master=supervisor_master,
        saver=saver,
        checkpoint_dir=checkpoint_path)

    # Start queue runners.
    coord = coordinator.Coordinator()
    threads = queue_runner.start_queue_runners(session, coord)

  with session:
    if not initialized:
      logging.warning('Failed to initialize from %s.', checkpoint_path)
      # TODO(ipolosukhin): This should be failing, but old code relies on that.
      session.run(variables.global_variables_initializer())
      if checkpoint_path:
        _restore_from_checkpoint(session, graph, checkpoint_path, saver)

    current_global_step = session.run(global_step_tensor)
    eval_results = None
    # TODO(amodei): Fix this to run through the eval set exactly once.
    step = 0
    eval_step = None
    feed_dict = None
    logging.info('Eval steps [%d,%s) for training step %d.', step,
                 'inf' if max_steps is None
                 else str(max_steps), current_global_step)
    try:
      try:
        while (max_steps is None) or (step < max_steps):
          step += 1
          start_time = time.time()
          feed_dict = feed_fn() if feed_fn is not None else None
          if update_op is not None:
            session.run(update_op, feed_dict=feed_dict)
          else:
            eval_results = session.run(eval_dict, feed_dict=feed_dict)
            eval_step = step

          # TODO(wicke): We should assert that the global step hasn't changed.
          if step % log_every_steps == 0:
            if eval_step is None or step != eval_step:
              eval_results = session.run(eval_dict, feed_dict=feed_dict)
              eval_step = step
            duration = time.time() - start_time
            logging.info('Results after %d steps (%.3f sec/batch): %s.',
                         step, float(duration),
                         _eval_results_to_str(eval_results))
      finally:
        if eval_results is None or step != eval_step:
          eval_results = session.run(eval_dict, feed_dict=feed_dict)
          eval_step = step
        # Stop session first, before queue runners.
        session.close()

        # Stop queue runners.
        try:
          coord.request_stop()
          coord.join(threads, stop_grace_period_secs=120)
        except (RuntimeError, errors.CancelledError) as e:
          logging.warning('Coordinator didn\'t stop cleanly: %s', e)

    # catch OutOfRangeError which is thrown when queue is out of data (and for
    # other reasons as well).
    except errors.OutOfRangeError as e:
      if max_steps is None:
        logging.info('Input queue is exhausted.')
      else:
        logging.warn('Input queue is exhausted: %s.', e)
    # catch StopIteration which is thrown is DataReader is out of data.
    except StopIteration as e:
      if max_steps is None:
        logging.info('Input iterator is exhausted.')
      else:
        logging.warn('Input iterator is exhausted: %s.', e)

  # Save summaries for this evaluation.
  _write_summary_results(output_dir, eval_results, current_global_step)

  return eval_results, current_global_step


def run_n(output_dict, feed_dict=None, restore_checkpoint_path=None, n=1):
  """Run `output_dict` tensors `n` times, with the same `feed_dict` each run.

  Args:
    output_dict: A `dict` mapping string names to tensors to run. Must all be
      from the same graph.
    feed_dict: `dict` of input values to feed each run.
    restore_checkpoint_path: A string containing the path to a checkpoint to
      restore.
    n: Number of times to repeat.

  Returns:
    A list of `n` `dict` objects, each containing values read from `output_dict`
    tensors.
  """
  return run_feeds(
      output_dict=output_dict,
      feed_dicts=itertools.repeat(feed_dict, n),
      restore_checkpoint_path=restore_checkpoint_path)


# TODO(ptucker): Add save_checkpoint_path.
def run_feeds_iter(output_dict, feed_dicts, restore_checkpoint_path=None):
  """Run `output_dict` tensors with each input in `feed_dicts`.

  If `restore_checkpoint_path` is supplied, restore from checkpoint. Otherwise,
  init all variables.

  Args:
    output_dict: A `dict` mapping string names to `Tensor` objects to run.
      Tensors must all be from the same graph.
    feed_dicts: Iterable of `dict` objects of input values to feed.
    restore_checkpoint_path: A string containing the path to a checkpoint to
      restore.

  Yields:
    A sequence of dicts of values read from `output_dict` tensors, one item
    yielded for each item in `feed_dicts`. Keys are the same as `output_dict`,
    values are the results read from the corresponding `Tensor` in
    `output_dict`.

  Raises:
    ValueError: if `output_dict` or `feed_dicts` is None or empty.
  """
  if not output_dict:
    raise ValueError('output_dict is invalid: %s.' % output_dict)
  if not feed_dicts:
    raise ValueError('feed_dicts is invalid: %s.' % feed_dicts)

  graph = contrib_ops.get_graph_from_inputs(output_dict.values())
  with graph.as_default() as g:
    with tf_session.Session('') as session:
      session.run(
          resources.initialize_resources(resources.shared_resources() +
                                         resources.local_resources()))
      if restore_checkpoint_path:
        _restore_from_checkpoint(session, g, restore_checkpoint_path)
      else:
        session.run(variables.global_variables_initializer())
      session.run(variables.local_variables_initializer())
      session.run(data_flow_ops.initialize_all_tables())
      coord = coordinator.Coordinator()
      threads = None
      try:
        threads = queue_runner.start_queue_runners(session, coord=coord)
        for f in feed_dicts:
          yield session.run(output_dict, f)
      finally:
        coord.request_stop()
        if threads:
          coord.join(threads, stop_grace_period_secs=120)


def run_feeds(*args, **kwargs):
  """See run_feeds_iter(). Returns a `list` instead of an iterator."""
  return list(run_feeds_iter(*args, **kwargs))


def infer(restore_checkpoint_path, output_dict, feed_dict=None):
  """Restore graph from `restore_checkpoint_path` and run `output_dict` tensors.

  If `restore_checkpoint_path` is supplied, restore from checkpoint. Otherwise,
  init all variables.

  Args:
    restore_checkpoint_path: A string containing the path to a checkpoint to
      restore.
    output_dict: A `dict` mapping string names to `Tensor` objects to run.
      Tensors must all be from the same graph.
    feed_dict: `dict` object mapping `Tensor` objects to input values to feed.

  Returns:
    Dict of values read from `output_dict` tensors. Keys are the same as
    `output_dict`, values are the results read from the corresponding `Tensor`
    in `output_dict`.

  Raises:
    ValueError: if `output_dict` or `feed_dicts` is None or empty.
  """
  return run_feeds(output_dict=output_dict,
                   feed_dicts=[feed_dict] if feed_dict is not None else [None],
                   restore_checkpoint_path=restore_checkpoint_path)[0]