Exemple #1
0
def bow_encoder(ids,
                vocab_size,
                embed_dim,
                sparse_lookup=True,
                initializer=None,
                regularizer=None,
                trainable=True,
                scope=None,
                reuse=None):
    """Maps a sequence of symbols to a vector per example by averaging embeddings.

  Args:
    ids: `[batch_size, doc_length]` `Tensor` or `SparseTensor` of type
      `int32` or `int64` with symbol ids.
    vocab_size: Integer number of symbols in vocabulary.
    embed_dim: Integer number of dimensions for embedding matrix.
    sparse_lookup: `bool`, if `True`, converts ids to a `SparseTensor`
        and performs a sparse embedding lookup. This is usually faster,
        but not desirable if padding tokens should have an embedding. Empty rows
        are assigned a special embedding.
    initializer: An initializer for the embeddings, if `None` default for
        current scope is used.
    regularizer: Optional regularizer for the embeddings.
    trainable: If `True` also add variables to the graph collection
      `GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).
    scope: Optional string specifying the variable scope for the op, required
        if `reuse=True`.
    reuse: If `True`, variables inside the op will be reused.

  Returns:
    Encoding `Tensor` `[batch_size, embed_dim]` produced by
    averaging embeddings.

  Raises:
    ValueError: If `embed_dim` or `vocab_size` are not specified.
  """
    if not vocab_size or not embed_dim:
        raise ValueError('Must specify vocab size and embedding dimension')
    with variable_scope.variable_scope(scope,
                                       'bow_encoder', [ids],
                                       reuse=reuse):
        embeddings = variables.model_variable('embeddings',
                                              shape=[vocab_size, embed_dim],
                                              initializer=initializer,
                                              regularizer=regularizer,
                                              trainable=trainable)
        if sparse_lookup:
            if isinstance(ids, ops.SparseTensor):
                sparse_ids = ids
            else:
                sparse_ids = sparse_ops.dense_to_sparse_tensor(ids)
            return contrib_embedding_ops.safe_embedding_lookup_sparse(
                [embeddings], sparse_ids, combiner='mean', default_id=0)
        else:
            if isinstance(ids, ops.SparseTensor):
                raise TypeError('ids are expected to be dense Tensor, got: %s',
                                ids)
            return math_ops.reduce_mean(embedding_ops.embedding_lookup(
                embeddings, ids),
                                        reduction_indices=1)
Exemple #2
0
def bow_encoder(ids,
                vocab_size,
                embed_dim,
                sparse_lookup=True,
                initializer=None,
                regularizer=None,
                trainable=True,
                scope=None,
                reuse=None):
  """Maps a sequence of symbols to a vector per example by averaging embeddings.

  Args:
    ids: `[batch_size, doc_length]` `Tensor` or `SparseTensor` of type
      `int32` or `int64` with symbol ids.
    vocab_size: Integer number of symbols in vocabulary.
    embed_dim: Integer number of dimensions for embedding matrix.
    sparse_lookup: `bool`, if `True`, converts ids to a `SparseTensor`
        and performs a sparse embedding lookup. This is usually faster,
        but not desirable if padding tokens should have an embedding. Empty rows
        are assigned a special embedding.
    initializer: An initializer for the embeddings, if `None` default for
        current scope is used.
    regularizer: Optional regularizer for the embeddings.
    trainable: If `True` also add variables to the graph collection
      `GraphKeys.TRAINABLE_VARIABLES` (see tf.Variable).
    scope: Optional string specifying the variable scope for the op, required
        if `reuse=True`.
    reuse: If `True`, variables inside the op will be reused.

  Returns:
    Encoding `Tensor` `[batch_size, embed_dim]` produced by
    averaging embeddings.

  Raises:
    ValueError: If `embed_dim` or `vocab_size` are not specified.
  """
  if not vocab_size or not embed_dim:
    raise ValueError('Must specify vocab size and embedding dimension')
  with variable_scope.variable_scope(
      scope, 'bow_encoder', [ids], reuse=reuse):
    embeddings = variables.model_variable(
        'embeddings', shape=[vocab_size, embed_dim],
        initializer=initializer, regularizer=regularizer,
        trainable=trainable)
    if sparse_lookup:
      if isinstance(ids, sparse_tensor.SparseTensor):
        sparse_ids = ids
      else:
        sparse_ids = sparse_ops.dense_to_sparse_tensor(ids)
      return contrib_embedding_ops.safe_embedding_lookup_sparse(
          [embeddings], sparse_ids, combiner='mean', default_id=0)
    else:
      if isinstance(ids, sparse_tensor.SparseTensor):
        raise TypeError('ids are expected to be dense Tensor, got: %s', ids)
      return math_ops.reduce_mean(
          embedding_ops.embedding_lookup(embeddings, ids),
          reduction_indices=1)
  def test_safe_embedding_lookup_sparse_3d_return_zero_vector(self):
    with self.cached_session():
      embedding_weights = self._random_weights()
      sparse_ids, sparse_weights = self._ids_and_weights_3d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, sparse_weights).eval())

      self.assertAllClose(embedding_lookup_result, [[
          (1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) / 3.0,
          [0] * 4, [0] * 4
      ], [embedding_weights[0][2], [0] * 4, [0] * 4]])
Exemple #4
0
  def test_safe_embedding_lookup_sparse_return_zero_vector(self):
    with self.test_session():
      embedding_weights = self._random_weights()
      sparse_ids, sparse_weights = self._ids_and_weights_2d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, sparse_weights).eval())

      self.assertAllClose(
          embedding_lookup_result,
          [(1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) /
           3.0, [0] * 4, [0] * 4, embedding_weights[0][2], [0] * 4])
  def test_safe_embedding_lookup_sparse_partitioned(self):
    with self.cached_session():
      embedding_weights = self._random_weights(num_shards=3)
      sparse_ids, _ = self._ids_and_weights_2d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, None).eval())

      embedding_weights = list(itertools.chain(*embedding_weights))
      self.assertAllClose(embedding_lookup_result,
                          [(embedding_weights[0] + embedding_weights[1]) / 2.0,
                           [0] * 4, [0] * 4, embedding_weights[2],
                           (embedding_weights[0] + embedding_weights[1]) / 2.0])
  def test_safe_embedding_lookup_sparse_no_weights(self):
    with self.cached_session():
      embedding_weights = self._random_weights()
      sparse_ids, _ = self._ids_and_weights_2d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, None).eval())

      self.assertAllClose(
          embedding_lookup_result,
          [(embedding_weights[0][0] + embedding_weights[0][1]) / 2.0, [0] * 4,
           [0] * 4, embedding_weights[0][2], (
               embedding_weights[0][0] + embedding_weights[0][1]) / 2.0])
  def test_safe_embedding_lookup_sparse_return_special_vector(self):
    with self.cached_session():
      embedding_weights = self._random_weights()
      sparse_ids, sparse_weights = self._ids_and_weights_2d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, sparse_weights, default_id=3).eval())

      self.assertAllClose(
          embedding_lookup_result,
          [(1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) /
           3.0, embedding_weights[0][3], embedding_weights[0][3],
           embedding_weights[0][2], embedding_weights[0][3]])
Exemple #8
0
  def test_safe_embedding_lookup_sparse_partitioned(self):
    with self.test_session():
      embedding_weights = self._random_weights(num_shards=3)
      sparse_ids, _ = self._ids_and_weights_2d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, None).eval())

      embedding_weights = list(itertools.chain(*embedding_weights))
      self.assertAllClose(embedding_lookup_result,
                          [(embedding_weights[0] + embedding_weights[1]) / 2.0,
                           [0] * 4, [0] * 4, embedding_weights[2],
                           (embedding_weights[0] + embedding_weights[1]) / 2.0])
Exemple #9
0
  def test_safe_embedding_lookup_sparse_no_weights(self):
    with self.test_session():
      embedding_weights = self._random_weights()
      sparse_ids, _ = self._ids_and_weights_2d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, None).eval())

      self.assertAllClose(
          embedding_lookup_result,
          [(embedding_weights[0][0] + embedding_weights[0][1]) / 2.0, [0] * 4,
           [0] * 4, embedding_weights[0][2],
           (embedding_weights[0][0] + embedding_weights[0][1]) / 2.0])
Exemple #10
0
  def test_safe_embedding_lookup_sparse_3d_return_special_vector(self):
    with self.test_session():
      embedding_weights = self._random_weights()
      sparse_ids, sparse_weights = self._ids_and_weights_3d()

      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
          embedding_weights, sparse_ids, sparse_weights, default_id=3).eval())

      self.assertAllClose(
          embedding_lookup_result,
          [[(1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) /
            3.0, embedding_weights[0][3], embedding_weights[0][3]], [
                embedding_weights[0][2], embedding_weights[0][3],
                embedding_weights[0][3]
            ]])
def _create_joint_embedding_lookup(columns_to_tensors,
                                   embedding_lookup_arguments,
                                   num_outputs,
                                   trainable,
                                   weight_collections):
  """Creates an embedding lookup for all columns sharing a single weight."""
  for arg in embedding_lookup_arguments:
    assert arg.weight_tensor is None, (
        'Joint sums for weighted sparse columns are not supported. '
        'Please use weighted_sum_from_feature_columns instead.')
    assert arg.combiner == 'sum', (
        'Combiners other than sum are not supported for joint sums. '
        'Please use weighted_sum_from_feature_columns instead.')
  assert len(embedding_lookup_arguments) >= 1, (
      'At least one column must be in the model.')
  prev_size = 0
  sparse_tensors = []
  for a in embedding_lookup_arguments:
    t = a.input_tensor
    values = t.values + prev_size
    prev_size += a.vocab_size
    sparse_tensors.append(
        ops.SparseTensor(t.indices,
                         values,
                         t.shape))
  sparse_tensor = sparse_ops.sparse_concat(1, sparse_tensors)
  with variable_scope.variable_scope(
      None, default_name='linear_weights', values=columns_to_tensors.values()):
    variable = contrib_variables.model_variable(
        name='weights',
        shape=[prev_size, num_outputs],
        dtype=dtypes.float32,
        initializer=init_ops.zeros_initializer,
        trainable=trainable,
        collections=weight_collections)
    if isinstance(variable, variables.Variable):
      variable = [variable]
    else:
      variable = variable._get_variable_list()  # pylint: disable=protected-access
    predictions = embedding_ops.safe_embedding_lookup_sparse(
        variable,
        sparse_tensor,
        sparse_weights=None,
        default_id=0,
        combiner='sum',
        name='_weights')
    return variable, predictions
Exemple #12
0
def _create_joint_embedding_lookup(columns_to_tensors,
                                   embedding_lookup_arguments,
                                   num_outputs,
                                   trainable,
                                   weight_collections):
  """Creates an embedding lookup for all columns sharing a single weight."""
  for arg in embedding_lookup_arguments:
    assert arg.weight_tensor is None, (
        'Joint sums for weighted sparse columns are not supported. '
        'Please use weighted_sum_from_feature_columns instead.')
    assert arg.combiner == 'sum', (
        'Combiners other than sum are not supported for joint sums. '
        'Please use weighted_sum_from_feature_columns instead.')
  assert len(embedding_lookup_arguments) >= 1, (
      'At least one column must be in the model.')
  prev_size = 0
  sparse_tensors = []
  for a in embedding_lookup_arguments:
    t = a.input_tensor
    values = t.values + prev_size
    prev_size += a.vocab_size
    sparse_tensors.append(
        ops.SparseTensor(t.indices,
                         values,
                         t.shape))
  sparse_tensor = sparse_ops.sparse_concat(1, sparse_tensors)
  with variable_scope.variable_scope(
      None, default_name='linear_weights', values=columns_to_tensors.values()):
    variable = contrib_variables.model_variable(
        name='weights',
        shape=[prev_size, num_outputs],
        dtype=dtypes.float32,
        initializer=init_ops.zeros_initializer,
        trainable=trainable,
        collections=weight_collections)
    if isinstance(variable, variables.Variable):
      variable = [variable]
    else:
      variable = variable._get_variable_list()  # pylint: disable=protected-access
    predictions = embedding_ops.safe_embedding_lookup_sparse(
        variable,
        sparse_tensor,
        sparse_weights=None,
        default_id=0,
        combiner='sum',
        name='_weights')
    return variable, predictions
def _create_embedding_lookup(column,
                             columns_to_tensors,
                             embedding_lookup_arguments,
                             num_outputs,
                             trainable,
                             weight_collections):
  """Creates variables and returns predictions for linear weights in a model.

  Args:
   column: the column we're working on.
   columns_to_tensors: a map from column name to tensors.
   embedding_lookup_arguments: arguments for embedding lookup.
   num_outputs: how many outputs.
   trainable: whether the variable we create is trainable.
   weight_collections: weights will be placed here.

  Returns:
  variables: the created embeddings.
  predictions: the computed predictions.
  """
  with variable_scope.variable_scope(
      None, default_name=column.name, values=columns_to_tensors.values()):
    variable = contrib_variables.model_variable(
        name='weights',
        shape=[embedding_lookup_arguments.vocab_size, num_outputs],
        dtype=dtypes.float32,
        initializer=embedding_lookup_arguments.initializer,
        trainable=trainable,
        collections=weight_collections)
    if isinstance(variable, variables.Variable):
      variable = [variable]
    else:
      variable = variable._get_variable_list()  # pylint: disable=protected-access
    predictions = embedding_ops.safe_embedding_lookup_sparse(
        variable,
        embedding_lookup_arguments.input_tensor,
        sparse_weights=embedding_lookup_arguments.weight_tensor,
        default_id=0,
        combiner=embedding_lookup_arguments.combiner,
        name=column.name + '_weights')
    return variable, predictions
Exemple #14
0
def _create_embedding_lookup(column,
                             columns_to_tensors,
                             embedding_lookup_arguments,
                             num_outputs,
                             trainable,
                             weight_collections):
  """Creates variables and returns predictions for linear weights in a model.

  Args:
   column: the column we're working on.
   columns_to_tensors: a map from column name to tensors.
   embedding_lookup_arguments: arguments for embedding lookup.
   num_outputs: how many outputs.
   trainable: whether the variable we create is trainable.
   weight_collections: weights will be placed here.

  Returns:
  variables: the created embeddings.
  predictions: the computed predictions.
  """
  with variable_scope.variable_scope(
      None, default_name=column.name, values=columns_to_tensors.values()):
    variable = contrib_variables.model_variable(
        name='weights',
        shape=[embedding_lookup_arguments.vocab_size, num_outputs],
        dtype=dtypes.float32,
        initializer=embedding_lookup_arguments.initializer,
        trainable=trainable,
        collections=weight_collections)
    if isinstance(variable, variables.Variable):
      variable = [variable]
    else:
      variable = variable._get_variable_list()  # pylint: disable=protected-access
    predictions = embedding_ops.safe_embedding_lookup_sparse(
        variable,
        embedding_lookup_arguments.input_tensor,
        sparse_weights=embedding_lookup_arguments.weight_tensor,
        default_id=0,
        combiner=embedding_lookup_arguments.combiner,
        name=column.name + '_weights')
    return variable, predictions
def _embeddings_from_arguments(column,
                               args,
                               weight_collections,
                               trainable,
                               output_rank=2):
  """Returns embeddings for a column based on the computed arguments.

  Args:
   column: the column name.
   args: the _DeepEmbeddingLookupArguments for this column.
   weight_collections: collections to store weights in.
   trainable: whether these embeddings should be trainable.
   output_rank: the desired rank of the returned `Tensor`. Inner dimensions will
     be combined to produce the desired rank.

  Returns:
   the embeddings.

  Raises:
   ValueError: if not possible to create.
  """
  # pylint: disable=protected-access
  input_tensor = layers._inner_flatten(args.input_tensor, output_rank)
  weight_tensor = None
  if args.weight_tensor is not None:
    weight_tensor = layers._inner_flatten(args.weight_tensor, output_rank)
  # pylint: enable=protected-access

  # This option is only enabled for scattered_embedding_column.
  if args.hash_key:
    embeddings = contrib_variables.model_variable(
        name='weights',
        shape=[args.vocab_size],
        dtype=dtypes.float32,
        initializer=args.initializer,
        trainable=trainable,
        collections=weight_collections)

    return embedding_ops.scattered_embedding_lookup_sparse(
        embeddings, input_tensor, args.dimension,
        hash_key=args.hash_key,
        combiner=args.combiner, name='lookup')

  if args.shared_embedding_name is not None:
    shared_embedding_collection_name = (
        'SHARED_EMBEDDING_COLLECTION_' + args.shared_embedding_name.upper())
    graph = ops.get_default_graph()
    shared_embedding_collection = (
        graph.get_collection_ref(shared_embedding_collection_name))
    shape = [args.vocab_size, args.dimension]
    if shared_embedding_collection:
      if len(shared_embedding_collection) > 1:
        raise ValueError('Collection %s can only contain one '
                         '(partitioned) variable.'
                         % shared_embedding_collection_name)
      else:
        embeddings = shared_embedding_collection[0]
        if embeddings.get_shape() != shape:
          raise ValueError('The embedding variable with name {} already '
                           'exists, but its shape does not match required '
                           'embedding shape  here. Please make sure to use '
                           'different shared_embedding_name for different '
                           'shared embeddings.'.format(
                               args.shared_embedding_name))
    else:
      embeddings = contrib_variables.model_variable(
          name=args.shared_embedding_name,
          shape=shape,
          dtype=dtypes.float32,
          initializer=args.initializer,
          trainable=trainable,
          collections=weight_collections)
      graph.add_to_collection(shared_embedding_collection_name, embeddings)
  else:
    embeddings = contrib_variables.model_variable(
        name='weights',
        shape=[args.vocab_size, args.dimension],
        dtype=dtypes.float32,
        initializer=args.initializer,
        trainable=trainable,
        collections=weight_collections)

  if isinstance(embeddings, variables.Variable):
    embeddings = [embeddings]
  else:
    embeddings = embeddings._get_variable_list()  # pylint: disable=protected-access
  # pylint: disable=protected-access
  _maybe_restore_from_checkpoint(
      column._checkpoint_path(), embeddings)
  return embedding_ops.safe_embedding_lookup_sparse(
      embeddings,
      input_tensor,
      sparse_weights=weight_tensor,
      combiner=args.combiner,
      name=column.name + 'weights',
      max_norm=args.max_norm)
Exemple #16
0
def _embeddings_from_arguments(column,
                               args,
                               weight_collections,
                               trainable,
                               output_rank=2):
  """Returns embeddings for a column based on the computed arguments.

  Args:
   column: the column name.
   args: the _DeepEmbeddingLookupArguments for this column.
   weight_collections: collections to store weights in.
   trainable: whether these embeddings should be trainable.
   output_rank: the desired rank of the returned `Tensor`. Inner dimensions will
     be combined to produce the desired rank.

  Returns:
   the embeddings.

  Raises:
   ValueError: if not possible to create.
  """
  # pylint: disable=protected-access
  input_tensor = layers._inner_flatten(args.input_tensor, output_rank)
  weight_tensor = None
  if args.weight_tensor is not None:
    weight_tensor = layers._inner_flatten(args.weight_tensor, output_rank)
  # pylint: enable=protected-access

  if args.hashed:
    embeddings = contrib_variables.model_variable(
        name='weights',
        shape=[args.vocab_size],
        dtype=dtypes.float32,
        initializer=args.initializer,
        trainable=trainable,
        collections=weight_collections)

    return embedding_ops.hashed_embedding_lookup_sparse(
        embeddings, input_tensor, args.dimension,
        combiner=args.combiner, name='lookup')

  if args.shared_embedding_name is not None:
    shared_embedding_collection_name = (
        'SHARED_EMBEDDING_COLLECTION_' + args.shared_embedding_name.upper())
    graph = ops.get_default_graph()
    shared_embedding_collection = (
        graph.get_collection_ref(shared_embedding_collection_name))
    shape = [args.vocab_size, args.dimension]
    if shared_embedding_collection:
      if len(shared_embedding_collection) > 1:
        raise ValueError('Collection %s can only contain one '
                         '(partitioned) variable.'
                         % shared_embedding_collection_name)
      else:
        embeddings = shared_embedding_collection[0]
        if embeddings.get_shape() != shape:
          raise ValueError('The embedding variable with name {} already '
                           'exists, but its shape does not match required '
                           'embedding shape  here. Please make sure to use '
                           'different shared_embedding_name for different '
                           'shared embeddings.'.format(
                               args.shared_embedding_name))
    else:
      embeddings = contrib_variables.model_variable(
          name=args.shared_embedding_name,
          shape=shape,
          dtype=dtypes.float32,
          initializer=args.initializer,
          trainable=trainable,
          collections=weight_collections)
      graph.add_to_collection(shared_embedding_collection_name, embeddings)
  else:
    embeddings = contrib_variables.model_variable(
        name='weights',
        shape=[args.vocab_size, args.dimension],
        dtype=dtypes.float32,
        initializer=args.initializer,
        trainable=trainable,
        collections=weight_collections)

  if isinstance(embeddings, variables.Variable):
    embeddings = [embeddings]
  else:
    embeddings = embeddings._get_variable_list()  # pylint: disable=protected-access
  # pylint: disable=protected-access
  _maybe_restore_from_checkpoint(
      column._checkpoint_path(), embeddings)
  return embedding_ops.safe_embedding_lookup_sparse(
      embeddings,
      input_tensor,
      sparse_weights=weight_tensor,
      combiner=args.combiner,
      name=column.name + 'weights',
      max_norm=args.max_norm)