🏭 Preprocessing Layers Factory

The PreprocessorLayerFactory class provides a convenient way to create and manage preprocessing layers for your machine learning models. It supports both standard Keras preprocessing layers and custom layers defined within the KDP framework.

🎡 Using Keras Preprocessing Layers

All preprocessing layers available in Keras can be used within the PreprocessorLayerFactory. You can access these layers by their class names. Here's an example of how to use a Keras preprocessing layer:

normalization_layer = PreprocessorLayerFactory.create_layer(
    "Normalization",
    axis=-1,
    mean=None,
    variance=None
)

Available layers:

• Normalization - Standardizes numerical features
• Discretization - Bins continuous features into discrete intervals
• CategoryEncoding - Converts categorical data into numeric representations
• Hashing - Performs feature hashing for categorical variables
• HashedCrossing - Creates feature crosses using hashing
• StringLookup - Converts string inputs to integer indices
• IntegerLookup - Maps integer inputs to indexed array positions
• TextVectorization - Processes raw text into encoded representations
• ... and more

🏗️ Custom KDP Preprocessing Layers

In addition to Keras layers, the PreprocessorLayerFactory includes several custom layers specific to the KDP framework. Here's a list of available custom layers:

cast_to_float32_layer(name='cast_to_float32', **kwargs) staticmethod

Create a CastToFloat32Layer layer.

Parameters:

Name	Type	Description	Default
name	str	The name of the layer.	'cast_to_float32'
**kwargs	dict	Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the CastToFloat32Layer layer.

create_layer(layer_class, name=None, **kwargs) staticmethod

Create a layer using the layer class name, automatically filtering kwargs based on the layer class.

Parameters:

Name	Type	Description	Default
layer_class	str | Class Object	The name of the layer class to be created (e.g., 'Normalization', 'Rescaling') or the class object itself.	required
name	str	The name of the layer. Optional.	None
**kwargs		Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the specified layer class.

date_encoding_layer(name='date_encoding_layer', **kwargs) staticmethod

Create a DateEncodingLayer layer.

Parameters:

Name	Type	Description	Default
name	str	The name of the layer.	'date_encoding_layer'
**kwargs	dict	Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the DateEncodingLayer layer.

date_parsing_layer(name='date_parsing_layer', **kwargs) staticmethod

Create a DateParsingLayer layer.

Parameters:

Name	Type	Description	Default
name	str	The name of the layer.	'date_parsing_layer'
**kwargs	dict	Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the DateParsingLayer layer.

date_season_layer(name='date_season_layer', **kwargs) staticmethod

Create a SeasonLayer layer.

Parameters:

Name	Type	Description	Default
name	str	The name of the layer.	'date_season_layer'
**kwargs	dict	Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the SeasonLayer layer.

distribution_aware_encoder(name='distribution_aware', num_bins=1000, epsilon=1e-06, detect_periodicity=True, handle_sparsity=True, adaptive_binning=True, mixture_components=3, prefered_distribution=None, **kwargs) staticmethod

Create a DistributionAwareEncoder layer.

Parameters:

Name	Type	Description	Default
name	str	Name of the layer	'distribution_aware'
num_bins	int	Number of bins for quantile encoding	1000
epsilon	float	Small value for numerical stability	1e-06
detect_periodicity	bool	Whether to detect and handle periodic patterns	True
handle_sparsity	bool	Whether to handle sparse data specially	True
adaptive_binning	bool	Whether to use adaptive binning	True
mixture_components	int	Number of components for mixture modeling	3
specified_distribution	DistributionType	Optional specific distribution type to use	required
**kwargs		Additional keyword arguments	{}

Returns:

Type	Description
tf.keras.layers.Layer	DistributionAwareEncoder layer

distribution_transform_layer(name='distribution_transform', transform_type='none', lambda_param=0.0, epsilon=1e-10, min_value=0.0, max_value=1.0, clip_values=True, auto_candidates=None, **kwargs) staticmethod

Create a DistributionTransformLayer layer.

Parameters:

Name	Type	Description	Default
name	str	Name of the layer	'distribution_transform'
transform_type	str	Type of transformation to apply	'none'
lambda_param	float	Parameter for parameterized transformations	0.0
epsilon	float	Small value for numerical stability	1e-10
min_value	float	Minimum value for min-max scaling	0.0
max_value	float	Maximum value for min-max scaling	1.0
clip_values	bool	Whether to clip values to the specified range	True
auto_candidates	list[str]	List of transformations to consider in auto mode	None
**kwargs		Additional keyword arguments	{}

Returns:

Type	Description
tf.keras.layers.Layer	DistributionTransformLayer layer

gated_linear_unit_layer(units, name='gated_linear_unit', **kwargs) staticmethod

Create a GatedLinearUnit layer.

Parameters:

Name	Type	Description	Default
units	int	Dimensionality of the output space	required
name	str	Name of the layer	'gated_linear_unit'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
GatedLinearUnit	tf.keras.layers.Layer	A GatedLinearUnit layer instance

gated_residual_network_layer(units, dropout_rate=0.2, name='gated_residual_network', **kwargs) staticmethod

Create a GatedResidualNetwork layer.

Parameters:

Name	Type	Description	Default
units	int	Dimensionality of the output space	required
dropout_rate	float	Fraction of the input units to drop	0.2
name	str	Name of the layer	'gated_residual_network'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
GatedResidualNetwork	tf.keras.layers.Layer	A GatedResidualNetwork layer instance

global_numerical_embedding_layer(global_embedding_dim=8, global_mlp_hidden_units=16, global_num_bins=10, global_init_min=-3.0, global_init_max=3.0, global_dropout_rate=0.1, global_use_batch_norm=True, global_pooling='average', name='global_numerical_embedding', **kwargs) staticmethod

Create a GlobalNumericalEmbedding layer.

Parameters:

Name	Type	Description	Default
global_embedding_dim	int	Dimension of the final global embedding	8
global_mlp_hidden_units	int	Number of hidden units in the global MLP	16
global_num_bins	int	Number of bins for discretization	10
global_init_min	float	Minimum value for initialization	-3.0
global_init_max	float	Maximum value for initialization	3.0
global_dropout_rate	float	Dropout rate for regularization	0.1
global_use_batch_norm	bool	Whether to use batch normalization	True
global_pooling	str	Pooling method to use ("average" or "max")	'average'
name	str	Name of the layer	'global_numerical_embedding'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
GlobalNumericalEmbedding	tf.keras.layers.Layer	A GlobalNumericalEmbedding layer instance

multi_resolution_attention_layer(num_heads, d_model, embedding_dim=32, name='multi_resolution_attention', **kwargs) staticmethod

Create a MultiResolutionTabularAttention layer.

Parameters:

Name	Type	Description	Default
num_heads	int	Number of attention heads	required
d_model	int	Dimensionality of the attention model	required
embedding_dim	int	Dimension for categorical embeddings	32
name	str	Name of the layer	'multi_resolution_attention'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
MultiResolutionTabularAttention	tf.keras.layers.Layer	A MultiResolutionTabularAttention layer instance

numerical_embedding_layer(embedding_dim=8, mlp_hidden_units=16, num_bins=10, init_min=-3.0, init_max=3.0, dropout_rate=0.1, use_batch_norm=True, name='numerical_embedding', **kwargs) staticmethod

Create a NumericalEmbedding layer.

Parameters:

Name	Type	Description	Default
embedding_dim	int	Dimension of the output embedding	8
mlp_hidden_units	int	Number of hidden units in the MLP	16
num_bins	int	Number of bins for discretization	10
init_min	float	Minimum value for initialization	-3.0
init_max	float	Maximum value for initialization	3.0
dropout_rate	float	Dropout rate for regularization	0.1
use_batch_norm	bool	Whether to use batch normalization	True
name	str	Name of the layer	'numerical_embedding'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
NumericalEmbedding	tf.keras.layers.Layer	A NumericalEmbedding layer instance

tabular_attention_layer(num_heads, d_model, name='tabular_attention', **kwargs) staticmethod

Create a TabularAttention layer.

Parameters:

Name	Type	Description	Default
num_heads	int	Number of attention heads	required
d_model	int	Dimensionality of the attention model	required
name	str	Name of the layer	'tabular_attention'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
TabularAttention	tf.keras.layers.Layer	A TabularAttention layer instance

text_preprocessing_layer(name='text_preprocessing', **kwargs) staticmethod

Create a TextPreprocessingLayer layer.

Parameters:

Name	Type	Description	Default
name	str	The name of the layer.	'text_preprocessing'
**kwargs	dict	Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the TextPreprocessingLayer layer.

transformer_block_layer(name='transformer', **kwargs) staticmethod

Create a TransformerBlock layer.

Parameters:

Name	Type	Description	Default
name	str	The name of the layer.	'transformer'
**kwargs	dict	Additional keyword arguments to pass to the layer constructor.	{}

Returns:

Type	Description
tf.keras.layers.Layer	An instance of the TransformerBlock layer.

variable_selection_layer(nr_features=None, units=16, dropout_rate=0.2, name='variable_selection', **kwargs) staticmethod

Create a VariableSelection layer.

Parameters:

Name	Type	Description	Default
nr_features	int	Number of input features	None
units	int	Dimensionality of the output space	16
dropout_rate	float	Fraction of the input units to drop	0.2
name	str	Name of the layer	'variable_selection'
**kwargs	dict	Additional arguments to pass to the layer	{}

Returns:

Name	Type	Description
VariableSelection	tf.keras.layers.Layer	A VariableSelection layer instance