scimba_torch.neural_nets.coordinates_based_nets.siren

Siren architecture implementation.

Classes

SirenLayer(in_size, out_size[, w0, c, ...])	Class representing a Siren Layer.
SirenNet(in_size, out_size, **kwargs)	Class representing a Siren architecture with optional ResNet layers.

class SirenLayer(in_size, out_size, w0=1, c=6, is_first=False, use_bias=True)

Bases: Module

Class representing a Siren Layer.

Parameters:

• in_size (int) – Dimension of the inputs.

• out_size (int) – Dimension of the outputs.

• w0 (int) – Frequency parameter. Defaults to 1.

• c (int) – Initialization parameter. Defaults to 6.

• is_first (bool) – Whether this is the first layer. Defaults to False.

• use_bias (bool) – Whether to use bias. Defaults to True.

in_size

Dimension of the inputs.

is_first

Whether this is the first layer.

out_size

Dimension of the outputs.

layer

The linear layer applied to the vector of features.

activation

The sine activation function.

init_(weight, bias, c, w0)

Init the weights of the layer using the specific Siren initialization.

Parameters:

• weight (Tensor) – The weight of the layer to initialize.

• bias (Tensor) – The bias of the layer to initialize.

• c (int) – A parameter for the weight initialization.

• w0 (int) – The frequency of the sinus activation function.

forward(x)

Apply the network to the inputs.

Parameters:

x (Tensor) – Input tensor.

Return type:

Tensor

Returns:

The result of the layer.

class SirenNet(in_size, out_size, **kwargs)

Bases: ScimbaModule

Class representing a Siren architecture with optional ResNet layers.

Parameters:

• in_size (int) – dimension of inputs

• out_size (int) – dimension of outputs

• **kwargs –

  Additional keyword arguments:

  • w (int): frequency for the internal layers’ activation functions

  • w0 (int): frequency for the first layer’s activation function

  • layer_sizes (list[int]): list of the size for each layer

  • use_res_net (bool, default=False): whether to use a ResNet architecture

layers

list of Siren layers, potentially with residual connections

forward(x, with_last_layer=True)

Apply the network to the inputs x.

Parameters:

• x (Tensor) – input tensor

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

Return type:

Tensor

Returns:

the result of the network