scimba_torch.neural_nets.coordinates_based_nets.pirate_net¶

PirateNet architecture implementation.

Classes

`PirateNet`([in_size, out_size, nb_features, ...])	A PirateNet neural network implementation.
`PirateNetBlock`([dim])	Implements a block of the PirateNet.

class PirateNetBlock(dim=1, **kwargs)[source]¶

Bases: ScimbaModule

Implements a block of the PirateNet.

Each block applies three linear transformations with activation to compute weighting matrices U and V, then updates the input x by combining these matrices using a residual scheme.

Parameters:

dim (int) – Input and output dimension of the block, default is 1
**kwargs – Additional parameters for block configuration

W_f¶: Linear layer for the f_l transformation

W_g¶: Linear layer for the g_l transformation

W_h¶: Linear layer for the h_l transformation

alpha¶: Trainable parameter for mixing the old and new value of x

activation¶: Activation function used in the block

forward(x, u, v)[source]¶

Applies the block transformation to the input x.

Parameters:

x (Tensor) – Input of the block
u (Tensor) – Weighting matrix U
v (Tensor) – Weighting matrix V

Return type:

Tensor

Returns:

Output of the block after transformation

class PirateNet(in_size=1, out_size=1, nb_features=1, nb_blocks=1, **kwargs)[source]¶

Bases: ScimbaModule

A PirateNet neural network implementation.

Parameters:

in_size (int) – Input dimension, default is 1
out_size (int) – Output dimension, default is 1
nb_features (int) – Number of features used for encoding, default is 1
nb_blocks (int) – Number of stacked PiranteNet_block layers, default is 1
**kwargs – Additional parameters for network configuration

in_size¶: Input dimension

out_size¶: Output dimension

nb_blocks¶: Number of residual blocks in the network

dim_hidden¶: Dimension of the latent space after encoding

embedding¶: Input encoding network

embedding_1¶: Linear layer to compute U

embedding_2¶: Linear layer to compute V

activation¶: Main activation function

blocks¶: list of PiranteNet_block blocks

output_layer¶: Output layer

activation_output¶: Final activation function applied to the output

forward(inputs, with_last_layer=True)[source]¶

Applies the network transformation to the inputs.

Parameters:

inputs (Tensor) – Input of the network
with_last_layer (bool) – If True, applies the output layer and final activation, default is True

Return type:

Tensor

Returns:

Output of the network after transformation