Mlp

Multi-Layer Perceptron (MLP) architectures.

nabla.nn.architectures.mlp.create_mlp_forward_and_loss(activation='relu')

Create a combined forward pass and loss computation function.

This function factory creates the forward_and_loss function needed for VJP computation in training loops.

Parameters:

activation (str) – Activation function for hidden layers

Returns:

Function that takes inputs and returns loss

Return type:

Callable

nabla.nn.architectures.mlp.create_mlp_config(layers, activation='relu', final_activation=None, init_method='he_normal', seed=42)

Create MLP configuration dictionary.

Parameters:

• layers (list[int]) – List of layer sizes [input, hidden1, hidden2, …, output]

• activation (str) – Activation function for hidden layers

• final_activation (str | None) – Optional activation for final layer

• init_method (str) – Weight initialization method

• seed (int) – Random seed for reproducibility

Returns:

Configuration dictionary with params and forward function

Return type:

dict

class nabla.nn.architectures.mlp.MLPBuilder

Bases: object

Builder class for creating MLP configurations.

__init__()

with_layers(layers)

Set layer sizes.

with_activation(activation)

Set hidden layer activation function.

with_final_activation(activation)

Set final layer activation function.

with_init_method(method)

Set weight initialization method.

with_seed(seed)

Set random seed.

build()

Build the MLP configuration.