data_pipeline.py

# Copyright 2018 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.
# ==============================================================================
"""Asynchronous data producer for the NCF pipeline."""

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

import atexit
import functools
import os
import sys
import tempfile
import threading
import time
import timeit
import traceback
import typing

import numpy as np
import six
from six.moves import queue
import tensorflow as tf
from absl import logging

import movielens
import constants as rconst
import popen_helper
import stat_utils


SUMMARY_TEMPLATE = """General:
{spacer}Num users: {num_users}
{spacer}Num items: {num_items}

Training:
{spacer}Positive count:          {train_pos_ct}
{spacer}Batch size:              {train_batch_size} {multiplier}
{spacer}Batch count per epoch:   {train_batch_ct}

Eval:
{spacer}Positive count:          {eval_pos_ct}
{spacer}Batch size:              {eval_batch_size} {multiplier}
{spacer}Batch count per epoch:   {eval_batch_ct}"""


class DatasetManager(object):
  """Helper class for handling TensorFlow specific data tasks.

  This class takes the (relatively) framework agnostic work done by the data
  constructor classes and handles the TensorFlow specific portions (TFRecord
  management, tf.Dataset creation, etc.).
  """

  def __init__(self,
               is_training,
               stream_files,
               batches_per_epoch,
               shard_root=None,
               deterministic=False,
               num_train_epochs=None):
    # type: (bool, bool, int, typing.Optional[str], bool, int) -> None
    """Constructs a `DatasetManager` instance.
    Args:
      is_training: Boolean of whether the data provided is training or
        evaluation data. This determines whether to reuse the data
        (if is_training=False) and the exact structure to use when storing and
        yielding data.
      stream_files: Boolean indicating whether data should be serialized and
        written to file shards.
      batches_per_epoch: The number of batches in a single epoch.
      shard_root: The base directory to be used when stream_files=True.
      deterministic: Forgo non-deterministic speedups. (i.e. sloppy=True)
      num_train_epochs: Number of epochs to generate. If None, then each
        call to `get_dataset()` increments the number of epochs requested.
    """
    self._is_training = is_training
    self._deterministic = deterministic
    self._stream_files = stream_files
    self._writers = []
    self._write_locks = [threading.RLock() for _ in
                         range(rconst.NUM_FILE_SHARDS)] if stream_files else []
    self._batches_per_epoch = batches_per_epoch
    self._epochs_completed = 0
    self._epochs_requested = num_train_epochs if num_train_epochs else 0
    self._shard_root = shard_root

    self._result_queue = queue.Queue()
    self._result_reuse = []

  @property
  def current_data_root(self):
    subdir = (rconst.TRAIN_FOLDER_TEMPLATE.format(self._epochs_completed)
              if self._is_training else rconst.EVAL_FOLDER)
    return os.path.join(self._shard_root, subdir)

  def buffer_reached(self):
    # Only applicable for training.
    return (self._epochs_completed - self._epochs_requested >=
            rconst.CYCLES_TO_BUFFER and self._is_training)

  @staticmethod
  def serialize(data):
    """Convert NumPy arrays into a TFRecords entry."""

    def create_int_feature(values):
      return tf.train.Feature(int64_list=tf.train.Int64List(value=list(values)))

    feature_dict = {
        k: create_int_feature(v.astype(np.int64)) for k, v in data.items()
    }

    return tf.train.Example(
        features=tf.train.Features(feature=feature_dict)).SerializeToString()

  @staticmethod
  def deserialize(serialized_data, batch_size=None, is_training=True):
    """Convert serialized TFRecords into tensors.

    Args:
      serialized_data: A tensor containing serialized records.
      batch_size: The data arrives pre-batched, so batch size is needed to
        deserialize the data.
      is_training: Boolean, whether data to deserialize to training data
        or evaluation data.
    """

    def _get_feature_map(batch_size, is_training=True):
      """Returns data format of the serialized tf record file."""

      if is_training:
        return {
            movielens.USER_COLUMN:
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64),
            movielens.ITEM_COLUMN:
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64),
            rconst.VALID_POINT_MASK:
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64),
            "labels":
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64)
        }
      else:
        return {
            movielens.USER_COLUMN:
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64),
            movielens.ITEM_COLUMN:
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64),
            rconst.DUPLICATE_MASK:
                tf.io.FixedLenFeature([batch_size], dtype=tf.int64)
        }

    features = tf.io.parse_single_example(
        serialized_data, _get_feature_map(batch_size, is_training=is_training))
    users = tf.reshape(
        tf.cast(features[movielens.USER_COLUMN], rconst.USER_DTYPE),
        (batch_size,))
    items = tf.reshape(
        tf.cast(features[movielens.ITEM_COLUMN], rconst.ITEM_DTYPE),
        (batch_size,))

    if is_training:
      valid_point_mask = tf.reshape(
          tf.cast(features[movielens.ITEM_COLUMN], tf.bool), (batch_size,))
      fake_dup_mask = tf.zeros_like(features[movielens.USER_COLUMN])
      return {
          movielens.USER_COLUMN: users,
          movielens.ITEM_COLUMN: items,
          rconst.VALID_POINT_MASK: valid_point_mask,
          rconst.TRAIN_LABEL_KEY:
              tf.reshape(tf.cast(features["labels"], tf.bool),
                         (batch_size, 1)),
          rconst.DUPLICATE_MASK: fake_dup_mask
      }
    else:
      labels = tf.reshape(
          tf.cast(tf.zeros_like(features[movielens.USER_COLUMN]), tf.bool),
          (batch_size, 1))
      fake_valid_pt_mask = tf.cast(
          tf.zeros_like(features[movielens.USER_COLUMN]), tf.bool)
      return {
          movielens.USER_COLUMN:
              users,
          movielens.ITEM_COLUMN:
              items,
          rconst.DUPLICATE_MASK:
              tf.reshape(
                  tf.cast(features[rconst.DUPLICATE_MASK], tf.bool),
                  (batch_size,)),
          rconst.VALID_POINT_MASK:
              fake_valid_pt_mask,
          rconst.TRAIN_LABEL_KEY:
              labels
      }

  def put(self, index, data):
    # type: (int, dict) -> None
    """Store data for later consumption.

    Because there are several paths for storing and yielding data (queues,
    lists, files) the data producer simply provides the data in a standard
    format at which point the dataset manager handles storing it in the correct
    form.

    Args:
      index: Used to select shards when writing to files.
      data: A dict of the data to be stored. This method mutates data, and
        therefore expects to be the only consumer.
    """
    if self._is_training:
      mask_start_index = data.pop(rconst.MASK_START_INDEX)
      batch_size = data[movielens.ITEM_COLUMN].shape[0]
      data[rconst.VALID_POINT_MASK] = np.less(
          np.arange(batch_size), mask_start_index)

    if self._stream_files:
      example_bytes = self.serialize(data)
      with self._write_locks[index % rconst.NUM_FILE_SHARDS]:
        self._writers[index % rconst.NUM_FILE_SHARDS].write(example_bytes)

    else:
      self._result_queue.put((
          data, data.pop("labels")) if self._is_training else data)

  def start_construction(self):
    if self._stream_files:
      tf.io.gfile.makedirs(self.current_data_root)
      template = os.path.join(self.current_data_root, rconst.SHARD_TEMPLATE)
      self._writers = [tf.io.TFRecordWriter(template.format(i))
                       for i in range(rconst.NUM_FILE_SHARDS)]

  def end_construction(self):
    if self._stream_files:
      [writer.close() for writer in self._writers]
      self._writers = []
      self._result_queue.put(self.current_data_root)

    self._epochs_completed += 1

  def data_generator(self, epochs_between_evals):
    """Yields examples during local training."""
    assert not self._stream_files
    assert self._is_training or epochs_between_evals == 1

    if self._is_training:
      for _ in range(self._batches_per_epoch * epochs_between_evals):
        yield self._result_queue.get(timeout=300)

    else:
      if self._result_reuse:
        assert len(self._result_reuse) == self._batches_per_epoch

        for i in self._result_reuse:
          yield i
      else:
        # First epoch.
        for _ in range(self._batches_per_epoch * epochs_between_evals):
          result = self._result_queue.get(timeout=300)
          self._result_reuse.append(result)
          yield result

  def increment_request_epoch(self):
    self._epochs_requested += 1

  def get_dataset(self, batch_size, epochs_between_evals):
    """Construct the dataset to be used for training and eval.

    For local training, data is provided through Dataset.from_generator. For
    remote training (TPUs) the data is first serialized to files and then sent
    to the TPU through a StreamingFilesDataset.

    Args:
      batch_size: The per-replica batch size of the dataset.
      epochs_between_evals: How many epochs worth of data to yield.
        (Generator mode only.)
    """
    self.increment_request_epoch()
    if self._stream_files:
      if epochs_between_evals > 1:
        raise ValueError("epochs_between_evals > 1 not supported for file "
                         "based dataset.")
      epoch_data_dir = self._result_queue.get(timeout=300)
      if not self._is_training:
        self._result_queue.put(epoch_data_dir)  # Eval data is reused.

      file_pattern = os.path.join(
          epoch_data_dir, rconst.SHARD_TEMPLATE.format("*"))
      # TODO(seemuch): remove this contrib import
      # pylint: disable=line-too-long
      from tensorflow.contrib.tpu.python.tpu.datasets import StreamingFilesDataset
      # pylint: enable=line-too-long
      dataset = StreamingFilesDataset(
          files=file_pattern, worker_job=popen_helper.worker_job(),
          num_parallel_reads=rconst.NUM_FILE_SHARDS, num_epochs=1,
          sloppy=not self._deterministic)
      map_fn = functools.partial(
          self.deserialize,
          batch_size=batch_size,
          is_training=self._is_training)
      dataset = dataset.map(map_fn, num_parallel_calls=16)

    else:
      types = {movielens.USER_COLUMN: rconst.USER_DTYPE,
               movielens.ITEM_COLUMN: rconst.ITEM_DTYPE}
      shapes = {movielens.USER_COLUMN: tf.TensorShape([batch_size]),
                movielens.ITEM_COLUMN: tf.TensorShape([batch_size])}

      if self._is_training:
        types[rconst.VALID_POINT_MASK] = np.bool
        shapes[rconst.VALID_POINT_MASK] = tf.TensorShape([batch_size])

        types = (types, np.bool)
        shapes = (shapes, tf.TensorShape([batch_size]))

      else:
        types[rconst.DUPLICATE_MASK] = np.bool
        shapes[rconst.DUPLICATE_MASK] = tf.TensorShape([batch_size])

      data_generator = functools.partial(
          self.data_generator, epochs_between_evals=epochs_between_evals)
      dataset = tf.data.Dataset.from_generator(
          generator=data_generator, output_types=types,
          output_shapes=shapes)

    return dataset.prefetch(16)

  def make_input_fn(self, batch_size):
    """Create an input_fn which checks for batch size consistency."""

    def input_fn(params):
      """Returns batches for training."""

      # Estimator passes batch_size during training and eval_batch_size during
      # eval. TPUEstimator only passes batch_size.
      param_batch_size = (params["batch_size"] if self._is_training else
                          params.get("eval_batch_size") or params["batch_size"])
      if batch_size != param_batch_size:
        raise ValueError("producer batch size ({}) differs from params batch "
                         "size ({})".format(batch_size, param_batch_size))

      epochs_between_evals = (params.get("epochs_between_evals", 1)
                              if self._is_training else 1)
      return self.get_dataset(batch_size=batch_size,
                              epochs_between_evals=epochs_between_evals)

    return input_fn


class BaseDataConstructor(threading.Thread):
  """Data constructor base class.

  This class manages the control flow for constructing data. It is not meant
  to be used directly, but instead subclasses should implement the following
  two methods:

    self.construct_lookup_variables
    self.lookup_negative_items

  """

  def __init__(
      self,
      maximum_number_epochs,  # type: int
      num_users,  # type: int
      num_items,  # type: int
      user_map,  # type: dict
      item_map,  # type: dict
      train_pos_users,  # type: np.ndarray
      train_pos_items,  # type: np.ndarray
      train_batch_size,  # type: int
      batches_per_train_step,  # type: int
      num_train_negatives,  # type: int
      eval_pos_users,  # type: np.ndarray
      eval_pos_items,  # type: np.ndarray
      eval_batch_size,  # type: int
      batches_per_eval_step,  # type: int
      stream_files,  # type: bool
      deterministic=False,  # type: bool
      epoch_dir=None,  # type: str
      num_train_epochs=None,  # type: int
      create_data_offline=False  # type: bool
  ):
    # General constants
    self._maximum_number_epochs = maximum_number_epochs
    self._num_users = num_users
    self._num_items = num_items
    self.user_map = user_map
    self.item_map = item_map
    self._train_pos_users = train_pos_users
    self._train_pos_items = train_pos_items
    self.train_batch_size = train_batch_size
    self._num_train_negatives = num_train_negatives
    self._batches_per_train_step = batches_per_train_step
    self._eval_pos_users = eval_pos_users
    self._eval_pos_items = eval_pos_items
    self.eval_batch_size = eval_batch_size
    self.num_train_epochs = num_train_epochs
    self.create_data_offline = create_data_offline

    # Training
    if self._train_pos_users.shape != self._train_pos_items.shape:
      raise ValueError(
          "User positives ({}) is different from item positives ({})".format(
              self._train_pos_users.shape, self._train_pos_items.shape))

    (self._train_pos_count,) = self._train_pos_users.shape
    self._elements_in_epoch = (1 + num_train_negatives) * self._train_pos_count
    self.train_batches_per_epoch = self._count_batches(
        self._elements_in_epoch, train_batch_size, batches_per_train_step)

    # Evaluation
    if eval_batch_size % (1 + rconst.NUM_EVAL_NEGATIVES):
      raise ValueError("Eval batch size {} is not divisible by {}".format(
          eval_batch_size, 1 + rconst.NUM_EVAL_NEGATIVES))
    self._eval_users_per_batch = int(
        eval_batch_size // (1 + rconst.NUM_EVAL_NEGATIVES))
    self._eval_elements_in_epoch = num_users * (1 + rconst.NUM_EVAL_NEGATIVES)
    self.eval_batches_per_epoch = self._count_batches(
        self._eval_elements_in_epoch, eval_batch_size, batches_per_eval_step)

    # Intermediate artifacts
    self._current_epoch_order = np.empty(shape=(0,))
    self._shuffle_iterator = None

    self._shuffle_with_forkpool = not stream_files
    if stream_files:
      self._shard_root = epoch_dir or tempfile.mkdtemp(prefix="ncf_")
      atexit.register(tf.io.gfile.rmtree, dirname=self._shard_root)
    else:
      self._shard_root = None

    self._train_dataset = DatasetManager(True, stream_files,
                                         self.train_batches_per_epoch,
                                         self._shard_root, deterministic,
                                         num_train_epochs)
    self._eval_dataset = DatasetManager(False, stream_files,
                                        self.eval_batches_per_epoch,
                                        self._shard_root, deterministic,
                                        num_train_epochs)

    # Threading details
    super(BaseDataConstructor, self).__init__()
    self.daemon = True
    self._stop_loop = False
    self._fatal_exception = None
    self.deterministic = deterministic

  def __str__(self):
    multiplier = ("(x{} devices)".format(self._batches_per_train_step)
                  if self._batches_per_train_step > 1 else "")
    summary = SUMMARY_TEMPLATE.format(
        spacer="  ", num_users=self._num_users, num_items=self._num_items,
        train_pos_ct=self._train_pos_count,
        train_batch_size=self.train_batch_size,
        train_batch_ct=self.train_batches_per_epoch,
        eval_pos_ct=self._num_users, eval_batch_size=self.eval_batch_size,
        eval_batch_ct=self.eval_batches_per_epoch, multiplier=multiplier)
    return super(BaseDataConstructor, self).__str__() + "\n" + summary

  @staticmethod
  def _count_batches(example_count, batch_size, batches_per_step):
    """Determine the number of batches, rounding up to fill all devices."""
    x = (example_count + batch_size - 1) // batch_size
    return (x + batches_per_step - 1) // batches_per_step * batches_per_step

  def stop_loop(self):
    self._stop_loop = True

  def construct_lookup_variables(self):
    """Perform any one time pre-compute work."""
    raise NotImplementedError

  def lookup_negative_items(self, **kwargs):
    """Randomly sample negative items for given users."""
    raise NotImplementedError

  def _run(self):
    atexit.register(self.stop_loop)
    self._start_shuffle_iterator()
    self.construct_lookup_variables()
    self._construct_training_epoch()
    self._construct_eval_epoch()
    for _ in range(self._maximum_number_epochs - 1):
      self._construct_training_epoch()
    self.stop_loop()

  def run(self):
    try:
      self._run()
    except Exception as e:
      # The Thread base class swallows stack traces, so unfortunately it is
      # necessary to catch and re-raise to get debug output
      traceback.print_exc()
      self._fatal_exception = e
      sys.stderr.flush()
      raise

  def _start_shuffle_iterator(self):
    if self._shuffle_with_forkpool:
      pool = popen_helper.get_forkpool(3, closing=False)
    else:
      pool = popen_helper.get_threadpool(1, closing=False)
    atexit.register(pool.close)
    args = [(self._elements_in_epoch, stat_utils.random_int32())
            for _ in range(self._maximum_number_epochs)]
    imap = pool.imap if self.deterministic else pool.imap_unordered
    self._shuffle_iterator = imap(stat_utils.permutation, args)

  def _get_training_batch(self, i):
    """Construct a single batch of training data.

    Args:
      i: The index of the batch. This is used when stream_files=True to assign
        data to file shards.
    """
    batch_indices = self._current_epoch_order[i * self.train_batch_size:
                                              (i + 1) * self.train_batch_size]
    (mask_start_index,) = batch_indices.shape

    batch_ind_mod = np.mod(batch_indices, self._train_pos_count)
    users = self._train_pos_users[batch_ind_mod]

    negative_indices = np.greater_equal(batch_indices, self._train_pos_count)
    negative_users = users[negative_indices]

    negative_items = self.lookup_negative_items(negative_users=negative_users)

    items = self._train_pos_items[batch_ind_mod]
    items[negative_indices] = negative_items

    labels = np.logical_not(negative_indices)

    # Pad last partial batch
    pad_length = self.train_batch_size - mask_start_index
    if pad_length:
      # We pad with arange rather than zeros because the network will still
      # compute logits for padded examples, and padding with zeros would create
      # a very "hot" embedding key which can have performance implications.
      user_pad = np.arange(pad_length, dtype=users.dtype) % self._num_users
      item_pad = np.arange(pad_length, dtype=items.dtype) % self._num_items
      label_pad = np.zeros(shape=(pad_length,), dtype=labels.dtype)
      users = np.concatenate([users, user_pad])
      items = np.concatenate([items, item_pad])
      labels = np.concatenate([labels, label_pad])

    self._train_dataset.put(i, {
        movielens.USER_COLUMN: users,
        movielens.ITEM_COLUMN: items,
        rconst.MASK_START_INDEX: np.array(mask_start_index, dtype=np.int32),
        "labels": labels,
    })

  def _wait_to_construct_train_epoch(self):
    count = 0
    while self._train_dataset.buffer_reached() and not self._stop_loop:
      time.sleep(0.01)
      count += 1
      if count >= 100 and np.log10(count) == np.round(np.log10(count)):
        logging.info(
            "Waited {} times for training data to be consumed".format(count))

  def _construct_training_epoch(self):
    """Loop to construct a batch of training data."""
    if not self.create_data_offline:
      self._wait_to_construct_train_epoch()

    start_time = timeit.default_timer()
    if self._stop_loop:
      return

    self._train_dataset.start_construction()
    map_args = list(range(self.train_batches_per_epoch))
    self._current_epoch_order = next(self._shuffle_iterator)

    get_pool = (popen_helper.get_fauxpool if self.deterministic else
                popen_helper.get_threadpool)
    with get_pool(6) as pool:
      pool.map(self._get_training_batch, map_args)
    self._train_dataset.end_construction()

    logging.info("Epoch construction complete. Time: {:.1f} seconds".format(
        timeit.default_timer() - start_time))

  @staticmethod
  def _assemble_eval_batch(users, positive_items, negative_items,
                           users_per_batch):
    """Construct duplicate_mask and structure data accordingly.

    The positive items should be last so that they lose ties. However, they
    should not be masked out if the true eval positive happens to be
    selected as a negative. So instead, the positive is placed in the first
    position, and then switched with the last element after the duplicate
    mask has been computed.

    Args:
      users: An array of users in a batch. (should be identical along axis 1)
      positive_items: An array (batch_size x 1) of positive item indices.
      negative_items: An array of negative item indices.
      users_per_batch: How many users should be in the batch. This is passed
        as an argument so that ncf_test.py can use this method.

    Returns:
      User, item, and duplicate_mask arrays.
    """
    items = np.concatenate([positive_items, negative_items], axis=1)

    # We pad the users and items here so that the duplicate mask calculation
    # will include padding. The metric function relies on all padded elements
    # except the positive being marked as duplicate to mask out padded points.
    if users.shape[0] < users_per_batch:
      pad_rows = users_per_batch - users.shape[0]
      padding = np.zeros(shape=(pad_rows, users.shape[1]), dtype=np.int32)
      users = np.concatenate([users, padding.astype(users.dtype)], axis=0)
      items = np.concatenate([items, padding.astype(items.dtype)], axis=0)

    duplicate_mask = stat_utils.mask_duplicates(items, axis=1).astype(np.bool)

    items[:, (0, -1)] = items[:, (-1, 0)]
    duplicate_mask[:, (0, -1)] = duplicate_mask[:, (-1, 0)]

    assert users.shape == items.shape == duplicate_mask.shape
    return users, items, duplicate_mask

  def _get_eval_batch(self, i):
    """Construct a single batch of evaluation data.

    Args:
      i: The index of the batch.
    """
    low_index = i * self._eval_users_per_batch
    high_index = (i + 1) * self._eval_users_per_batch
    users = np.repeat(self._eval_pos_users[low_index:high_index, np.newaxis],
                      1 + rconst.NUM_EVAL_NEGATIVES, axis=1)
    positive_items = self._eval_pos_items[low_index:high_index, np.newaxis]
    negative_items = (self.lookup_negative_items(negative_users=users[:, :-1])
                      .reshape(-1, rconst.NUM_EVAL_NEGATIVES))

    users, items, duplicate_mask = self._assemble_eval_batch(
        users, positive_items, negative_items, self._eval_users_per_batch)

    self._eval_dataset.put(i, {
        movielens.USER_COLUMN: users.flatten(),
        movielens.ITEM_COLUMN: items.flatten(),
        rconst.DUPLICATE_MASK: duplicate_mask.flatten(),
    })

  def _construct_eval_epoch(self):
    """Loop to construct data for evaluation."""
    if self._stop_loop:
      return

    start_time = timeit.default_timer()

    self._eval_dataset.start_construction()
    map_args = [i for i in range(self.eval_batches_per_epoch)]

    get_pool = (popen_helper.get_fauxpool if self.deterministic else
                popen_helper.get_threadpool)
    with get_pool(6) as pool:
      pool.map(self._get_eval_batch, map_args)
    self._eval_dataset.end_construction()

    logging.info("Eval construction complete. Time: {:.1f} seconds".format(
        timeit.default_timer() - start_time))

  def make_input_fn(self, is_training):
    # It isn't feasible to provide a foolproof check, so this is designed to
    # catch most failures rather than provide an exhaustive guard.
    if self._fatal_exception is not None:
      raise ValueError("Fatal exception in the data production loop: {}"
                       .format(self._fatal_exception))

    return (
        self._train_dataset.make_input_fn(self.train_batch_size) if is_training
        else self._eval_dataset.make_input_fn(self.eval_batch_size))

  def increment_request_epoch(self):
    self._train_dataset.increment_request_epoch()


class DummyConstructor(threading.Thread):
  """Class for running with synthetic data."""

  def __init__(self, *args, **kwargs):
    super(DummyConstructor, self).__init__(*args, **kwargs)
    self.train_batches_per_epoch = rconst.SYNTHETIC_BATCHES_PER_EPOCH
    self.eval_batches_per_epoch = rconst.SYNTHETIC_BATCHES_PER_EPOCH

  def run(self):
    pass

  def stop_loop(self):
    pass

  def increment_request_epoch(self):
    pass

  @staticmethod
  def make_input_fn(is_training):
    """Construct training input_fn that uses synthetic data."""

    def input_fn(params):
      """Returns dummy input batches for training."""

      # Estimator passes batch_size during training and eval_batch_size during
      # eval. TPUEstimator only passes batch_size.
      batch_size = (params["batch_size"] if is_training else
                    params.get("eval_batch_size") or params["batch_size"])
      num_users = params["num_users"]
      num_items = params["num_items"]

      users = tf.random.uniform([batch_size], dtype=tf.int32, minval=0,
                                maxval=num_users)
      items = tf.random.uniform([batch_size], dtype=tf.int32, minval=0,
                                maxval=num_items)

      if is_training:
        valid_point_mask = tf.cast(tf.random.uniform(
            [batch_size], dtype=tf.int32, minval=0, maxval=2), tf.bool)
        labels = tf.cast(tf.random.uniform(
            [batch_size], dtype=tf.int32, minval=0, maxval=2), tf.bool)
        data = {
            movielens.USER_COLUMN: users,
            movielens.ITEM_COLUMN: items,
            rconst.VALID_POINT_MASK: valid_point_mask,
        }, labels
      else:
        dupe_mask = tf.cast(tf.random.uniform([batch_size], dtype=tf.int32,
                                              minval=0, maxval=2), tf.bool)
        data = {
            movielens.USER_COLUMN: users,
            movielens.ITEM_COLUMN: items,
            rconst.DUPLICATE_MASK: dupe_mask,
        }

      dataset = tf.data.Dataset.from_tensors(data).repeat(
          rconst.SYNTHETIC_BATCHES_PER_EPOCH * params["batches_per_step"])
      dataset = dataset.prefetch(32)
      return dataset

    return input_fn


class MaterializedDataConstructor(BaseDataConstructor):
  """Materialize a table of negative examples for fast negative generation.

  This class creates a table (num_users x num_items) containing all of the
  negative examples for each user. This table is conceptually ragged; that is to
  say the items dimension will have a number of unused elements at the end equal
  to the number of positive elements for a given user. For instance:

  num_users = 3
  num_items = 5
  positives = [[1, 3], [0], [1, 2, 3, 4]]

  will generate a negative table:
  [
    [0         2         4         int32max  int32max],
    [1         2         3         4         int32max],
    [0         int32max  int32max  int32max  int32max],
  ]

  and a vector of per-user negative counts, which in this case would be:
    [3, 4, 1]

  When sampling negatives, integers are (nearly) uniformly selected from the
  range [0, per_user_neg_count[user]) which gives a column_index, at which
  point the negative can be selected as:
    negative_table[user, column_index]

  This technique will not scale; however MovieLens is small enough that even
  a pre-compute which is quadratic in problem size will still fit in memory. A
  more scalable lookup method is in the works.
  """
  def __init__(self, *args, **kwargs):
    super(MaterializedDataConstructor, self).__init__(*args, **kwargs)
    self._negative_table = None
    self._per_user_neg_count = None

  def construct_lookup_variables(self):
    # Materialize negatives for fast lookup sampling.
    start_time = timeit.default_timer()
    inner_bounds = np.argwhere(self._train_pos_users[1:] -
                               self._train_pos_users[:-1])[:, 0] + 1
    (upper_bound,) = self._train_pos_users.shape
    index_bounds = [0] + inner_bounds.tolist() + [upper_bound]
    self._negative_table = np.zeros(shape=(self._num_users, self._num_items),
                                    dtype=rconst.ITEM_DTYPE)

    # Set the table to the max value to make sure the embedding lookup will fail
    # if we go out of bounds, rather than just overloading item zero.
    self._negative_table += np.iinfo(rconst.ITEM_DTYPE).max
    assert self._num_items < np.iinfo(rconst.ITEM_DTYPE).max

    # Reuse arange during generation. np.delete will make a copy.
    full_set = np.arange(self._num_items, dtype=rconst.ITEM_DTYPE)

    self._per_user_neg_count = np.zeros(
        shape=(self._num_users,), dtype=np.int32)

    # Threading does not improve this loop. For some reason, the np.delete
    # call does not parallelize well. Multiprocessing incurs too much
    # serialization overhead to be worthwhile.
    for i in range(self._num_users):
      positives = self._train_pos_items[index_bounds[i]:index_bounds[i+1]]
      negatives = np.delete(full_set, positives)
      self._per_user_neg_count[i] = self._num_items - positives.shape[0]
      self._negative_table[i, :self._per_user_neg_count[i]] = negatives

    logging.info("Negative sample table built. Time: {:.1f} seconds".format(
        timeit.default_timer() - start_time))

  def lookup_negative_items(self, negative_users, **kwargs):
    negative_item_choice = stat_utils.very_slightly_biased_randint(
        self._per_user_neg_count[negative_users])
    return self._negative_table[negative_users, negative_item_choice]


class BisectionDataConstructor(BaseDataConstructor):
  """Use bisection to index within positive examples.

  This class tallies the number of negative items which appear before each
  positive item for a user. This means that in order to select the ith negative
  item for a user, it only needs to determine which two positive items bound
  it at which point the item id for the ith negative is a simply algebraic
  expression.
  """
  def __init__(self, *args, **kwargs):
    super(BisectionDataConstructor, self).__init__(*args, **kwargs)
    self.index_bounds = None
    self._sorted_train_pos_items = None
    self._total_negatives = None

  def _index_segment(self, user):
    lower, upper = self.index_bounds[user:user+2]
    items = self._sorted_train_pos_items[lower:upper]

    negatives_since_last_positive = np.concatenate(
        [items[0][np.newaxis], items[1:] - items[:-1] - 1])

    return np.cumsum(negatives_since_last_positive)

  def construct_lookup_variables(self):
    start_time = timeit.default_timer()
    inner_bounds = np.argwhere(self._train_pos_users[1:] -
                               self._train_pos_users[:-1])[:, 0] + 1
    (upper_bound,) = self._train_pos_users.shape
    self.index_bounds = np.array([0] + inner_bounds.tolist() + [upper_bound])

    # Later logic will assume that the users are in sequential ascending order.
    assert np.array_equal(self._train_pos_users[self.index_bounds[:-1]],
                          np.arange(self._num_users))

    self._sorted_train_pos_items = self._train_pos_items.copy()

    for i in range(self._num_users):
      lower, upper = self.index_bounds[i:i+2]
      self._sorted_train_pos_items[lower:upper].sort()

    self._total_negatives = np.concatenate([
        self._index_segment(i) for i in range(self._num_users)])

    logging.info("Negative total vector built. Time: {:.1f} seconds".format(
        timeit.default_timer() - start_time))

  def lookup_negative_items(self, negative_users, **kwargs):
    output = np.zeros(shape=negative_users.shape, dtype=rconst.ITEM_DTYPE) - 1

    left_index = self.index_bounds[negative_users]
    right_index = self.index_bounds[negative_users + 1] - 1

    num_positives = right_index - left_index + 1
    num_negatives = self._num_items - num_positives
    neg_item_choice = stat_utils.very_slightly_biased_randint(num_negatives)

    # Shortcuts:
    # For points where the negative is greater than or equal to the tally before
    # the last positive point there is no need to bisect. Instead the item id
    # corresponding to the negative item choice is simply:
    #   last_postive_index + 1 + (neg_choice - last_negative_tally)
    # Similarly, if the selection is less than the tally at the first positive
    # then the item_id is simply the selection.
    #
    # Because MovieLens organizes popular movies into low integers (which is
    # preserved through the preprocessing), the first shortcut is very
    # efficient, allowing ~60% of samples to bypass the bisection. For the same
    # reason, the second shortcut is rarely triggered (<0.02%) and is therefore
    # not worth implementing.
    use_shortcut = neg_item_choice >= self._total_negatives[right_index]
    output[use_shortcut] = (
        self._sorted_train_pos_items[right_index] + 1 +
        (neg_item_choice - self._total_negatives[right_index])
    )[use_shortcut]

    if np.all(use_shortcut):
      # The bisection code is ill-posed when there are no elements.
      return output

    not_use_shortcut = np.logical_not(use_shortcut)
    left_index = left_index[not_use_shortcut]
    right_index = right_index[not_use_shortcut]
    neg_item_choice = neg_item_choice[not_use_shortcut]

    num_loops = np.max(
        np.ceil(np.log2(num_positives[not_use_shortcut])).astype(np.int32))

    for i in range(num_loops):
      mid_index = (left_index + right_index) // 2
      right_criteria = self._total_negatives[mid_index] > neg_item_choice
      left_criteria = np.logical_not(right_criteria)

      right_index[right_criteria] = mid_index[right_criteria]
      left_index[left_criteria] = mid_index[left_criteria]

    # Expected state after bisection pass:
    #   The right index is the smallest index whose tally is greater than the
    #   negative item choice index.

    assert np.all((right_index - left_index) <= 1)

    output[not_use_shortcut] = (
        self._sorted_train_pos_items[right_index] -
        (self._total_negatives[right_index] - neg_item_choice)
    )

    assert np.all(output >= 0)

    return output


def get_constructor(name):
  if name == "bisection":
    return BisectionDataConstructor
  if name == "materialized":
    return MaterializedDataConstructor
  raise ValueError("Unrecognized constructor: {}".format(name))