model.CTRNN

Parameters

• dims (list, default: [1, 100, 1]):
  The dimensions of the network in the format [input_dim, hidden_size, output_dim].

• preact_noise (float, default: 0):
  Standard deviation of Gaussian noise added to the pre-activation state.

• postact_noise (float, default: 0):
  Standard deviation of Gaussian noise added to the post-activation state.

• activation (str, default: "relu"):
  Activation function for the recurrent layer. Options: "relu", "sigmoid", "tanh", "retanh".

• dt (float, default: 10):
  Time step size.

• tau (float, default: 100):
  Time constant for the recurrent dynamics.

• biases (list or single value, default: None):
  Specifies bias initialization for each layer.
  • Options for each layer: "zero", "normal", "uniform", or None.
  • Can also directly pass a numpy.ndarray or torch.Tensor.
• weights (list or single value, default: "uniform"):
  Specifies weight initialization for each layer.
  • Options for each layer: "normal", "uniform", or directly pass numpy.ndarray or torch.Tensor.
• sparsity_masks (list or None, default: None):
  Masks defining sparsity for each layer.
  • Each value can be None or a numpy.ndarray/torch.Tensor.
• ei_masks (list or None, default: None):
  Masks enforcing Dale’s law (excitatory/inhibitory constraints) for each layer.
  • Each value can be None or a numpy.ndarray/torch.Tensor.
• plasticity_masks (list or None, default: None):
  Masks controlling plasticity for each layer.
  • Each value can be None or a numpy.ndarray/torch.Tensor.
• synapse_growth_masks (list or None, default: None):
  Masks controlling synapse growth probability.
  • Each value can be None or a numpy.ndarray/torch.Tensor.

Methods

to(device)

Moves the network and its layers to the specified device (CPU or GPU).

Parameters:

• device (torch.device):
  Target device (e.g., torch.device("cuda") or "cpu").

Usage:

ctrnn.to(torch.device("cuda"))

forward(x, init_state=None)

Performs a forward pass through the network.

Parameters:

• x (torch.Tensor):
  Input tensor of shape (batch_size, n_timesteps, input_dim).
• init_state (torch.Tensor, optional, default: None):
  Initial state of the recurrent layer.

Returns:

• output (torch.Tensor):
  Output tensor of the network.
• hidden_states (dict):
  Dictionary containing hidden states.

Usage:

output, states = ctrnn.forward(input_tensor, init_state)

train()

Sets the network to training mode.
Restores pre-activation and post-activation noise values and resumes enforcing constraints.

Usage:

ctrnn.train()

eval()

Sets the network to evaluation mode.
Disables pre-activation and post-activation noise and pauses enforcing constraints.

Usage:

ctrnn.eval()

apply_plasticity()

Applies plasticity masks to the weight gradients of the network’s layers.

Usage:

ctrnn.apply_plasticity()

print_layer()

Prints the specifications of each layer in the network.

Usage:

ctrnn.print_layer()

plot_layers()

Plots the weight matrices and distributions of each layer in the network.

Usage:

ctrnn.plot_layers()