🏗️ Models API Reference

Welcome to the KMR Models documentation! All models are designed to work exclusively with Keras 3 and provide high-level abstractions for common tabular data processing tasks.

What You'll Find Here

Each model includes detailed documentation with: - ✨ Complete parameter descriptions with types and defaults - 🎯 Usage examples showing real-world applications - ⚡ Best practices and performance considerations - 🎨 When to use guidance for each model - 🔧 Implementation notes for developers

Ready-to-Use Models

These models provide complete architectures that you can use out-of-the-box or customize for your specific needs.

Base Classes

All models inherit from BaseModel ensuring consistent behavior and Keras 3 compatibility.

🏗️ Core Models

🚀 BaseFeedForwardModel

Flexible feed-forward model architecture for tabular data with customizable layers.

kmr.models.feed_forward.BaseFeedForwardModel

BaseFeedForwardModel(
    feature_names,
    hidden_units,
    output_units=1,
    dropout_rate=0.0,
    activation="relu",
    preprocessing_model=None,
    kernel_initializer="glorot_uniform",
    bias_initializer="zeros",
    kernel_regularizer=None,
    bias_regularizer=None,
    activity_regularizer=None,
    kernel_constraint=None,
    bias_constraint=None,
    **kwargs
)

Base feed forward neural network model.

This model implements a basic feed forward neural network with configurable hidden layers, activations, and regularization options.

Example

# Create a simple feed forward model
model = BaseFeedForwardModel(
    feature_names=['feature1', 'feature2'],
    hidden_units=[64, 32],
    output_units=1
)

# Compile and train the model
model.compile(optimizer='adam', loss='mse')
model.fit(train_dataset, epochs=10)

Initialize Feed Forward Neural Network.

Parameters:

Name	Type	Description	Default
feature_names	list[str]	list of feature names.	required
hidden_units	list[int]	list of hidden layer units.	required
output_units	int	Number of output units.	1
dropout_rate	float	Dropout rate.	0.0
activation	str	Activation function.	'relu'
preprocessing_model	Model | None	Optional preprocessing model.	None
kernel_initializer	str | Any | None	Weight initializer.	'glorot_uniform'
bias_initializer	str | Any | None	Bias initializer.	'zeros'
kernel_regularizer	str | Any | None	Weight regularizer.	None
bias_regularizer	str | Any | None	Bias regularizer.	None
activity_regularizer	str | Any | None	Activity regularizer.	None
kernel_constraint	str | Any | None	Weight constraint.	None
bias_constraint	str | Any | None	Bias constraint.	None
**kwargs	Any	Additional arguments.	{}

Functions

from_config classmethod

from_config(config)

Create model from configuration.

Parameters:

Name	Type	Description	Default
config	dict[str, Any]	Dict containing model configuration.	required

Returns:

Type	Description
BaseFeedForwardModel	Instantiated model.

🎯 Advanced Models

🧩 SFNEBlock

Sparse Feature Network Ensemble block for advanced feature processing and ensemble learning.

kmr.models.SFNEBlock

This module implements a SFNEBlock (Slow-Fast Neural Engine Block) model that combines slow and fast processing paths for feature extraction. It's a building block for the Terminator model.

Classes

SFNEBlock

SFNEBlock(
    input_dim,
    output_dim=None,
    hidden_dim=64,
    num_layers=2,
    slow_network_layers=3,
    slow_network_units=128,
    preprocessing_model=None,
    name=None,
    **kwargs
)

Slow-Fast Neural Engine Block for feature processing.

This model combines a slow network path and a fast processing path to extract features. It uses a SlowNetwork to generate hyper-kernels, which are then used by a HyperZZWOperator to compute context-dependent weights. These weights are further processed by global and local convolutions before being combined.

Parameters:

Name	Type	Description	Default
input_dim	int	Dimension of the input features.	required
output_dim	int	Dimension of the output features. Default is same as input_dim.	None
hidden_dim	int	Number of hidden units in the network. Default is 64.	64
num_layers	int	Number of layers in the network. Default is 2.	2
slow_network_layers	int	Number of layers in the slow network. Default is 3.	3
slow_network_units	int	Number of units per layer in the slow network. Default is 128.	128
preprocessing_model	Model | None	Optional preprocessing model to apply before the main processing.	None
name	str | None	Optional name for the model.	None

Input shape

2D tensor with shape: (batch_size, input_dim) or a dictionary with feature inputs

Output shape

2D tensor with shape: (batch_size, output_dim)

Example

import keras
from kmr.models import SFNEBlock

# Create sample input data
x = keras.random.normal((32, 16))  # 32 samples, 16 features

# Create the model
sfne = SFNEBlock(input_dim=16, output_dim=8)
y = sfne(x)
print("Output shape:", y.shape)  # (32, 8)

Initialize the SFNEBlock model.

Parameters:

Name	Type	Description	Default
input_dim	int	Input dimension.	required
output_dim	int	Output dimension.	None
hidden_dim	int	Hidden dimension.	64
num_layers	int	Number of layers.	2
slow_network_layers	int	Number of slow network layers.	3
slow_network_units	int	Number of units in slow network.	128
preprocessing_model	Model | None	Preprocessing model.	None
name	str | None	Name of the model.	None
**kwargs	Any	Additional keyword arguments.	{}

Functions

from_config classmethod

from_config(config)

Creates a model from its configuration.

Parameters:

Name	Type	Description	Default
config	dict[str, Any]	Dictionary containing the model configuration.	required

Returns:

Type	Description
SFNEBlock	A new instance of the model.

🎭 TerminatorModel

Comprehensive tabular model that combines multiple SFNE blocks for complex data tasks.

kmr.models.TerminatorModel

This module implements a TerminatorModel that combines multiple SFNE blocks for advanced feature processing. It's designed for complex tabular data modeling tasks.

Classes

TerminatorModel

TerminatorModel(
    input_dim,
    context_dim,
    output_dim,
    hidden_dim=64,
    num_layers=2,
    num_blocks=3,
    slow_network_layers=3,
    slow_network_units=128,
    preprocessing_model=None,
    name=None,
    **kwargs
)

Terminator model for advanced feature processing.

This model stacks multiple SFNE blocks to process features in a hierarchical manner. It's designed for complex tabular data modeling tasks where feature interactions are important.

Parameters:

Name	Type	Description	Default
input_dim	int	Dimension of the input features.	required
context_dim	int	Dimension of the context features.	required
output_dim	int	Dimension of the output.	required
hidden_dim	int	Number of hidden units in the network. Default is 64.	64
num_layers	int	Number of layers in the network. Default is 2.	2
num_blocks	int	Number of SFNE blocks to stack. Default is 3.	3
slow_network_layers	int	Number of layers in each slow network. Default is 3.	3
slow_network_units	int	Number of units per layer in each slow network. Default is 128.	128
preprocessing_model	Model | None	Optional preprocessing model to apply before the main processing.	None
name	str | None	Optional name for the model.	None

Input shape

List of 2D tensors with shapes: [(batch_size, input_dim), (batch_size, context_dim)]

Output shape

2D tensor with shape: (batch_size, output_dim)

Example

import keras
from kmr.models import TerminatorModel

# Create sample input data
x = keras.random.normal((32, 16))  # 32 samples, 16 features
context = keras.random.normal((32, 8))  # 32 samples, 8 context features

# Create the model
terminator = TerminatorModel(input_dim=16, context_dim=8, output_dim=1)
y = terminator([x, context])
print("Output shape:", y.shape)  # (32, 1)

Initialize the TerminatorModel.

Parameters:

Name	Type	Description	Default
input_dim	int	Input dimension.	required
context_dim	int	Context dimension.	required
output_dim	int	Output dimension.	required
hidden_dim	int	Hidden dimension.	64
num_layers	int	Number of layers.	2
num_blocks	int	Number of blocks.	3
slow_network_layers	int	Number of slow network layers.	3
slow_network_units	int	Number of units in slow network.	128
preprocessing_model	Model | None	Preprocessing model.	None
name	str | None	Name of the model.	None
**kwargs	Any	Additional keyword arguments.	{}

Functions

from_config classmethod

from_config(config)

Creates a model from its configuration.

Parameters:

Name	Type	Description	Default
config	dict[str, Any]	Dictionary containing the model configuration.	required

Returns:

Type	Description
TerminatorModel	A new instance of the model.

🔧 Base Classes

🏛️ BaseModel

Base class for all KMR models, providing common functionality and Keras 3 compatibility.

kmr.models._base.BaseModel

Functions

inspect_signatures

inspect_signatures(model)

Inspect the model signatures.

Parameters:

Name	Type	Description	Default
self		write your description	required
model	Model	write your description	required