v2.0

I've finally updated this, fixed, and added some modern functions developer use today. This includes new activations, more flexibility on your trainings, and most importantly speed and effectiveness. Dimanet revived?

New Activation Functions

ReLU (Rectified Linear Unit)

• dimanet_act_relu() - The standard ReLU function
• Perfect for hidden layers in deep networks
• Helps mitigate vanishing gradient problem

Tanh (Hyperbolic Tangent)

• dimanet_act_tanh() zero-centered activation
• Output range arrays: [-1, 1]
• Often performs better than sigmoid for hidden layers

Leaky ReLU

• dimanet_act_leaky_relu() - prevents the famous "dying relu" problem most of the frameworks have (thank you!)
• Configurable alpha parameter for negative inputs (finally)
• dimanet_set_leaky_relu_alpha() to customize slope

Enhanced API

Flexible Activation Configuration

// Set different activations for hidden and output layers
dimanet_set_activation_functions(ann, dimanet_act_relu, dimanet_act_linear);

// Configure Leaky ReLU parameters  
dimanet_set_leaky_relu_alpha(ann, 0.1);

What's all of that?

• ReLU: Deep networks, computer vision, most modern architectures
• Tanh: Recurrent networks, when zero-centered outputs are beneficial
• Leaky ReLU: Preventing dead neurons in large networks
• Sigmoid: Binary classification, output layers (remains default)