gnn_tracking.models.mlp

gnn_tracking.models.mlp#

Fully connected neural network implementations

Classes#

MLP

Multi Layer Perceptron, using ReLu as activation function.

ResFCNN

Fully connected NN with residual connections.

HeterogeneousResFCNN

Separate FCNNs for pixel and strip data, with residual connections.

ResMLP

Fully connected NN w/ residual connections and Gaussian init

Functions#

get_pixel_mask(→ torch.Tensor)

Module Contents#

class gnn_tracking.models.mlp.MLP(input_size: int, output_size: int, hidden_dim: int | None, L=3, *, bias=True, include_last_activation=False)#

Bases: torch.nn.Module

Multi Layer Perceptron, using ReLu as activation function.

Parameters:

  • input_size – Input feature dimension

  • output_size – Output feature dimension

  • hidden_dim – Feature dimension of the hidden layers. If None: Choose maximum of input/output size

  • L – Total number of layers (1 initial layer, L-2 hidden layers, 1 output layer)

  • bias – Include bias in linear layer?

  • include_last_activation – Include activation function for the last layer?

layers#
reset_parameters()#
forward(x)#
class gnn_tracking.models.mlp.ResFCNN(*, in_dim: int, hidden_dim: int, out_dim: int, depth: int, alpha: float = 0.6, bias: bool = True)#

Bases: torch.nn.Module

Fully connected NN with residual connections.

Parameters:

  • in_dim – Input dimension

  • hidden_dim – Hidden dimension

  • out_dim – Output dimension = embedding space

  • depth – 1 input encoder layer, depth-1 hidden layers, 1 output encoder layer

  • alpha – strength of the residual connection

_encoder#
_decoder#
_layers#
_alpha = 0.6#
static _reset_layer_parameters(layer, var: float)#
forward(x: torch.Tensor, **ignore) → torch.Tensor#
gnn_tracking.models.mlp.get_pixel_mask(layer: torch.Tensor) → torch.Tensor#
class gnn_tracking.models.mlp.HeterogeneousResFCNN(*, in_dim: int, out_dim: int, hidden_dim: int, depth: int, alpha: float = 0.6, bias: bool = True)#

Bases: torch.nn.Module

Separate FCNNs for pixel and strip data, with residual connections. For parameters, see ResFCNN.

pixel_fcnn#
strip_fcnn#
forward(x: torch.Tensor, layer: torch.Tensor) → torch.Tensor#
class gnn_tracking.models.mlp.ResMLP(in_dim: int, out_dim: int, hidden_dim: int, depth: int = 4, beta: float = 1.0, gamma_0: float = 1.0, eta_0: float = 0.01, activation: torch.nn.Module = Tanh, optimizer: str = 'adam', bias: bool = True, **kwargs)#

Bases: torch.nn.Module

Fully connected NN w/ residual connections and Gaussian init Args: in_dim: input dimension out_dim: output dimension width: # neurons per internal layer beta: strength of the residual connection gamma_0: tuning of final layer output normalisation depth: number of hidden layers

layers#
in_dim#
out_dim#
width#
beta = 1.0#
gamma_0 = 1.0#
eta_0 = 0.01#
gamma#
depth = 4#
act#
lr#
reset_parameters()#
get_lr(optimizer)#
forward(x)#
Contents