NetVecor

NetVector – Interpreter for flat Vectors as feedforward neural networks.

This module provides a lightweight helper class that interprets a flat Vector (e.g. created via dim_type=’net’) as a fully connected feedforward network with arbitrary layer structure and activation function.

NetVector does not contain any trainable parameters or evolutionary logic itself. Instead, it unpacks and applies a flat vector (weights + biases) to input data.

Typical use case:

• Use Vector as evolvable parameter container (mutation, crossover etc.)

• Use NetVector to define the network structure and perform forward evaluations

Example

para = Vector(…) # created via normal_initializer_net net = NetVector(dim=[1, 8, 1], activation=”tanh”) y = net.forward(x, para.vector)

class evolib.representation.netvector.NetVector(dim, activation='tanh')

Bases: object

classmethod from_config(cfg, module)

Return type:

NetVector

forward(x, vector)

Evaluates the network on input x using the flat parameter vector. Activation is applied after all but the last layer.

Parameters:

• x (np.ndarray) – Input vector (shape: [input_dim] or [batch, input_dim])

• vector (np.ndarray) – Flat parameter vector with correct dimension

Returns:

Output of the network

Return type:

np.ndarray