class UnivariateInteractionModel(InteractionModel):
"""
Defines an interaction model that configures feature layer operations
on individual features
"""
def __init__(
self,
feature_config: FeatureConfig,
feature_layer_keys_to_fns: dict,
tfrecord_type: str,
max_sequence_size: int = 0,
):
self.feature_config = feature_config
self.tfrecord_type = tfrecord_type
self.max_sequence_size = max_sequence_size
self.feature_layer_map = FeatureLayerMap()
self.feature_layer_map.add_fns(feature_layer_keys_to_fns)
def feature_layer_op(self, inputs: Dict[str, Input]):
if self.tfrecord_type == TFRecordTypeKey.EXAMPLE:
train_features, metadata_features = define_example_feature_layer(
feature_config=self.feature_config, feature_layer_map=self.feature_layer_map
)(inputs)
else:
train_features, metadata_features = define_sequence_example_feature_layer(
feature_config=self.feature_config,
feature_layer_map=self.feature_layer_map,
max_sequence_size=self.max_sequence_size,
)(inputs)
return train_features, metadata_features
def transform_features_op(self, train_features, metadata_features):
# TODO: Make train_features a dictionary
train_features = tf.concat(train_features, axis=-1, name="train_features")
return train_features, metadata_features
class UnivariateInteractionModel(InteractionModel):
"""
Defines an interaction model that configures feature layer operations
on individual features
"""
def __init__(
self,
feature_config: FeatureConfig,
tfrecord_type: str,
feature_layer_keys_to_fns: dict = {},
max_sequence_size: int = 0,
file_io: FileIO = None,
):
"""
Constructor for instantiating a UnivariateInteractionModel
Parameters
----------
feature_config : `FeatureConfig` object
FeatureConfig object that defines list of model features
and the feature transformation functions to be used on each
tfrecord_type : {"example", "sequence_example"}
Type of TFRecord protobuf being used for model training
feature_layer_keys_to_fns : dict
Dictionary of custom feature transformation functions to be applied
on the input features
max_sequence_size : int, optional
Maximum size of the sequence in SequenceExample protobuf
file_io : FileIO object
`FileIO` object that handles read write operations
"""
self.feature_config = feature_config
self.tfrecord_type = tfrecord_type
self.max_sequence_size = max_sequence_size
self.feature_layer_map = FeatureLayerMap()
self.feature_layer_map.add_fns(feature_layer_keys_to_fns)
self.file_io = file_io
def feature_layer_op(self, inputs: Dict[str, Input]):
"""
Convert input tensorflow features into numeric
train features and metadata features by applying respective feature
transformation functions as specified in the FeatureConfig
Parameters
----------
inputs : dict
Dictionary of the inputs to the tensorflow keras model
Returns
-------
train_features : dict
Dict of feature tensors that are used for training
metadata_features : dict
Dictionary of feature tensors that can be used for
computing custom metrics and losses
"""
train_features, metadata_features = define_feature_layer(
feature_config=self.feature_config,
tfrecord_type=self.tfrecord_type,
feature_layer_map=self.feature_layer_map,
file_io=self.file_io,
)(inputs)
return train_features, metadata_features
def transform_features_op(self, train_features: Dict[str, tf.Tensor],
metadata_features: Dict[str, tf.Tensor]):
"""
Convert train and metadata features which have feature layer
functions applied to them into dense numeric tensors.
Sorts the features by name and concats the individual features
into a dense tensor.
Parameters
----------
inputs : dict
Dictionary of the inputs to the tensorflow keras model
Returns
-------
train_features : `tf.Tensor`
Dense tensor object that is used for training
metadata_features : dict
Dictionary of feature tensors that can be used for
computing custom metrics and losses
"""
# Sorting the train features dictionary so that we control the order
train_features_list = [
train_features[k] for k in sorted(train_features)
]
# Concat all train features to get a dense feature vector
train_features_transformed = tf.concat(train_features_list,
axis=-1,
name="train_features")
return train_features_transformed, metadata_features
class UnivariateInteractionModel(InteractionModel):
"""
Defines an interaction model that configures feature layer operations
on individual features
"""
def __init__(
self,
feature_config: FeatureConfig,
feature_layer_keys_to_fns: dict,
tfrecord_type: str,
max_sequence_size: int = 0,
file_io: FileIO = None,
):
self.feature_config = feature_config
self.tfrecord_type = tfrecord_type
self.max_sequence_size = max_sequence_size
self.feature_layer_map = FeatureLayerMap()
self.feature_layer_map.add_fns(feature_layer_keys_to_fns)
self.file_io = file_io
def feature_layer_op(self, inputs: Dict[str, Input]):
"""
Apply feature layer functions on each of the tf.keras.Input
Args:
inputs: dictionary of keras input symbolic tensors
Returns:
train_features: dictionary of feature tensors that can be used for training
metadata_features: dictionary of feature tensors that can be used as additional metadata
"""
train_features, metadata_features = define_feature_layer(
feature_config=self.feature_config,
tfrecord_type=self.tfrecord_type,
feature_layer_map=self.feature_layer_map,
file_io=self.file_io,
)(inputs)
return train_features, metadata_features
def transform_features_op(self, train_features: Dict[str, tf.Tensor],
metadata_features: Dict[str, tf.Tensor]):
"""
Transform train_features and metadata_features after the
univariate feature_layer fns have been applied.
Args:
train_features: dictionary of feature tensors that can be used for training
metadata_features: dictionary of feature tensors that can be used as additional metadata
Returns:
train_features: single dense trainable feature tensor
metadata_features: dictionary of metadata feature tensors
"""
# Sorting the train features dictionary so that we control the order
train_features_list = [
train_features[k] for k in sorted(train_features)
]
# Concat all train features to get a dense feature vector
train_features_transformed = tf.concat(train_features_list,
axis=-1,
name="train_features")
return train_features_transformed, metadata_features