scimba_torch.neural_nets.coordinates_based_nets.discontinuous_mlp

A Multi-Layer Perceptron (MLP) with discontinuous layers.

Each hidden layer can either be discontinuous or regular.

Classes

DiscontinuousLayer(in_size, out_size, **kwargs)	Class that encodes a fully connected layer which can be discontinuous or not.
DiscontinuousMLP(in_size, out_size, **kwargs)	A Multi-Layer Perceptron (MLP) with discontinuous layers.

class DiscontinuousLayer(in_size, out_size, **kwargs)

Bases: Module

Class that encodes a fully connected layer which can be discontinuous or not.

It computes: \(y = \sigma(Ax + b) + \epsilon * H(Ax + b)\).

where \(H(x)\) is the Heaviside function and \(\epsilon\) is a learnable vector.

Parameters:

• in_size (int) – The input dimension size.

• out_size (int) – The output dimension size.

• **kwargs –

  Keyword arguments including:

  • activation_type (str): The activation function type. Defaults to “tanh”.

  • dis (bool): If True, the layer includes the discontinuous term, otherwise it behaves as a regular layer. Defaults to True.

Example

>>> layer = DiscontinuousLayer(10, 5, activation_type='relu', dis=True)

linearlayer

The linear transformation applied to the inputs.

eps

The parameters which multiply the Heaviside function. The size is the size of the output of the layer.

forward(inputs)

Apply the network to the inputs.

Parameters:

inputs (Tensor) – Input tensor

Return type:

Tensor

Returns:

The result of the network

class DiscontinuousMLP(in_size, out_size, **kwargs)

Bases: ScimbaModule

A Multi-Layer Perceptron (MLP) with discontinuous layers.

Each hidden layer can either be discontinuous or regular.

Parameters:

• in_size (int) – Input dimension.

• out_size (int) – Output dimension.

• **kwargs –

  Keyword arguments including:

  • activation_type (str): The type of activation function to be used for hidden layers. Defaults to “tanh”.

  • activation_output (str): The type of activation function for the output layer. Defaults to “id”.

  • layer_sizes (list[int]): List of sizes for each hidden layer. Defaults to [10, 20, 20, 20, 5].

  • layer_type (list[bool]): List of booleans indicating whether each hidden layer should be discontinuous. Defaults to [False, False, True, False, False].

Raises:

ValueError – If layer_sizes and layer_type lists have different lengths.

Example

>>> model = DiscontinuousMLP(
...     10, 5, activation_type="relu", activation_output="tanh",
...     layer_sizes=[50, 30], layer_type=[False, True, False]
... )

hidden_layers

The list of discontinuous or regular layers in the model.

output_layer

The final output layer.

forward(inputs, with_last_layer=True)

Forward pass through the discontinuous MLP network.

Parameters:

• inputs (Tensor) – Input tensor.

• with_last_layer (bool) – Whether to apply the final output layer. Defaults to True.

Return type:

Tensor

Returns:

Output tensor after processing through the MLP.