def _prepare_inputs_for_rnn(sequence_features, context_features,
sequence_feature_columns, num_unroll):
"""Prepares features batched by the SQSS for input to a state-saving RNN.
Args:
sequence_features: A dict of sequence feature name to `Tensor` or
`SparseTensor`, with `Tensor`s of shape `[batch_size, num_unroll, ...]`
or `SparseTensors` of dense shape `[batch_size, num_unroll, d]`.
context_features: A dict of context feature name to `Tensor`, with
tensors of shape `[batch_size, 1, ...]` and type float32.
sequence_feature_columns: An iterable containing all the feature columns
describing sequence features. All items in the set should be instances
of classes derived from `FeatureColumn`.
num_unroll: Python integer, how many time steps to unroll at a time.
The input sequences of length `k` are then split into `k / num_unroll`
many segments.
Returns:
features_by_time: A list of length `num_unroll` with `Tensor` entries of
shape `[batch_size, sum(sequence_features dimensions) +
sum(context_features dimensions)]` of type float32.
Context features are copied into each time step.
"""
def _tile(feature):
return array_ops.squeeze(
array_ops.tile(array_ops.expand_dims(feature, 1), [1, num_unroll, 1]),
axis=2)
for feature in sequence_features.values():
if isinstance(feature, sparse_tensor.SparseTensor):
# Explicitly set dense_shape's shape to 3 ([batch_size, num_unroll, d])
# since it can't be statically inferred.
feature.dense_shape.set_shape([3])
sequence_features = layers.sequence_input_from_feature_columns(
columns_to_tensors=sequence_features,
feature_columns=sequence_feature_columns,
weight_collections=None,
scope=None)
# Explicitly set shape along dimension 1 to num_unroll for the unstack op.
sequence_features.set_shape([None, num_unroll, None])
if not context_features:
return array_ops.unstack(sequence_features, axis=1)
# TODO(jtbates): Call layers.input_from_feature_columns for context features.
context_features = [
_tile(context_features[k]) for k in sorted(context_features)
]
return array_ops.unstack(
array_ops.concat(
[sequence_features, array_ops.stack(context_features, 2)], axis=2),
axis=1)
def _prepare_inputs_for_rnn(sequence_features, context_features,
sequence_feature_columns, num_unroll):
"""Prepares features batched by the SQSS for input to a state-saving RNN.
Args:
sequence_features: A dict of sequence feature name to `Tensor` or
`SparseTensor`, with `Tensor`s of shape `[batch_size, num_unroll, ...]`
or `SparseTensors` of dense shape `[batch_size, num_unroll, d]`.
context_features: A dict of context feature name to `Tensor`, with
tensors of shape `[batch_size, 1, ...]` and type float32.
sequence_feature_columns: An iterable containing all the feature columns
describing sequence features. All items in the set should be instances
of classes derived from `FeatureColumn`.
num_unroll: Python integer, how many time steps to unroll at a time.
The input sequences of length `k` are then split into `k / num_unroll`
many segments.
Returns:
features_by_time: A list of length `num_unroll` with `Tensor` entries of
shape `[batch_size, sum(sequence_features dimensions) +
sum(context_features dimensions)]` of type float32.
Context features are copied into each time step.
"""
def _tile(feature):
return array_ops.squeeze(array_ops.tile(
array_ops.expand_dims(feature, 1), [1, num_unroll, 1]),
axis=2)
for feature in sequence_features.values():
if isinstance(feature, sparse_tensor.SparseTensor):
# Explicitly set dense_shape's shape to 3 ([batch_size, num_unroll, d])
# since it can't be statically inferred.
feature.dense_shape.set_shape([3])
sequence_features = layers.sequence_input_from_feature_columns(
columns_to_tensors=sequence_features,
feature_columns=sequence_feature_columns,
weight_collections=None,
scope=None)
# Explicitly set shape along dimension 1 to num_unroll for the unstack op.
sequence_features.set_shape([None, num_unroll, None])
if not context_features:
return array_ops.unstack(sequence_features, axis=1)
# TODO(jtbates): Call layers.input_from_feature_columns for context features.
context_features = [
_tile(context_features[k]) for k in sorted(context_features)
]
return array_ops.unstack(array_ops.concat(
[sequence_features,
array_ops.stack(context_features, 2)], axis=2),
axis=1)
def build_sequence_input(features,
sequence_feature_columns,
context_feature_columns,
weight_collections=None,
scope=None):
"""Combine sequence and context features into input for an RNN.
Args:
features: A `dict` containing the input and (optionally) sequence length
information and initial state.
sequence_feature_columns: An iterable containing all the feature columns
describing sequence features. All items in the set should be instances
of classes derived from `FeatureColumn`.
context_feature_columns: An iterable containing all the feature columns
describing context features i.e. features that apply across all time
steps. All items in the set should be instances of classes derived from
`FeatureColumn`.
weight_collections: List of graph collections to which weights are added.
scope: Optional scope, passed through to parsing ops.
Returns:
A `Tensor` of dtype `float32` and shape `[batch_size, padded_length, ?]`.
This will be used as input to an RNN.
"""
features = features.copy()
features.update(
layers.transform_features(
features,
list(sequence_feature_columns) +
list(context_feature_columns or [])))
sequence_input = layers.sequence_input_from_feature_columns(
columns_to_tensors=features,
feature_columns=sequence_feature_columns,
weight_collections=weight_collections,
scope=scope)
if context_feature_columns is not None:
context_input = layers.input_from_feature_columns(
columns_to_tensors=features,
feature_columns=context_feature_columns,
weight_collections=weight_collections,
scope=scope)
sequence_input = _concatenate_context_input(sequence_input,
context_input)
return sequence_input
def _get_model_input(self, features, weight_collections=None, scope=None):
# TODO(jamieas): add option to use context to construct initial state rather
# than appending it to sequence input.
initial_state = features.get(self._initial_state_key)
sequence_input = layers.sequence_input_from_feature_columns(
columns_to_tensors=features,
feature_columns=self._sequence_feature_columns,
weight_collections=weight_collections,
scope=scope)
if self._context_feature_columns is not None:
context_input = layers.input_from_feature_columns(
columns_to_tensors=features,
feature_columns=self._context_feature_columns,
weight_collections=weight_collections,
scope=scope)
sequence_input = _concatenate_context_input(sequence_input, context_input)
return initial_state, sequence_input
def _get_model_input(self, features, weight_collections=None, scope=None):
# TODO(jamieas): add option to use context to construct initial state rather
# than appending it to sequence input.
initial_state = features.get(self._initial_state_key)
sequence_input = layers.sequence_input_from_feature_columns(
columns_to_tensors=features,
feature_columns=self._sequence_feature_columns,
weight_collections=weight_collections,
scope=scope)
if self._context_feature_columns is not None:
context_input = layers.input_from_feature_columns(
columns_to_tensors=features,
feature_columns=self._context_feature_columns,
weight_collections=weight_collections,
scope=scope)
sequence_input = _concatenate_context_input(sequence_input, context_input)
return initial_state, sequence_input
def build_sequence_input(features,
sequence_feature_columns,
context_feature_columns,
weight_collections=None,
scope=None):
"""Combine sequence and context features into input for an RNN.
Args:
features: A `dict` containing the input and (optionally) sequence length
information and initial state.
sequence_feature_columns: An iterable containing all the feature columns
describing sequence features. All items in the set should be instances
of classes derived from `FeatureColumn`.
context_feature_columns: An iterable containing all the feature columns
describing context features i.e. features that apply across all time
steps. All items in the set should be instances of classes derived from
`FeatureColumn`.
weight_collections: List of graph collections to which weights are added.
scope: Optional scope, passed through to parsing ops.
Returns:
A `Tensor` of dtype `float32` and shape `[batch_size, padded_length, ?]`.
This will be used as input to an RNN.
"""
features = features.copy()
features.update(layers.transform_features(
features,
list(sequence_feature_columns) + list(context_feature_columns or [])))
sequence_input = layers.sequence_input_from_feature_columns(
columns_to_tensors=features,
feature_columns=sequence_feature_columns,
weight_collections=weight_collections,
scope=scope)
if context_feature_columns is not None:
context_input = layers.input_from_feature_columns(
columns_to_tensors=features,
feature_columns=context_feature_columns,
weight_collections=weight_collections,
scope=scope)
sequence_input = _concatenate_context_input(sequence_input, context_input)
return sequence_input
