void NeuralNetwork::CalculateNeuronActivation(size_t i){
double x = m_neurons[i].m_activesum;
m_neurons[i].m_activesum = 0;
// Apply the activation function
double y = 0.0;
switch (m_neurons[i].m_activation_function_type)
{
case SIGNED_SIGMOID:
y = af_sigmoid_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_SIGMOID:
y = af_sigmoid_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case TANH:
y = af_tanh(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case TANH_CUBIC:
y = af_tanh_cubic(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case SIGNED_STEP:
y = af_step_signed(x, m_neurons[i].m_b);
break;
case UNSIGNED_STEP:
y = af_step_unsigned(x, m_neurons[i].m_b);
break;
case SIGNED_GAUSS:
y = af_gauss_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_GAUSS:
y = af_gauss_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case ABS:
y = af_abs(x, m_neurons[i].m_b);
break;
case SIGNED_SINE:
y = af_sine_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_SINE:
y = af_sine_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case SIGNED_SQUARE:
y = af_square_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_SQUARE:
y = af_square_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case LINEAR:
y = af_linear(x, m_neurons[i].m_b);
break;
default:
y = af_sigmoid_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
}
m_neurons[i].m_activation = y;
}
void NeuralNetwork::ActivateLeaky(double a_dtime)
{
// Loop connections. Calculate each connection's output signal.
for (unsigned int i = 0; i < m_connections.size(); i++)
{
m_connections[i].m_signal =
m_neurons[m_connections[i].m_source_neuron_idx].m_activation
* m_connections[i].m_weight;
}
// Loop the connections again. This time add the signals to the target neurons.
// This will largely require out of order memory writes. This is the one loop where
// this will happen.
for (unsigned int i = 0; i < m_connections.size(); i++)
{
m_neurons[m_connections[i].m_target_neuron_idx].m_activesum +=
m_connections[i].m_signal;
}
// Now we have the leaky integrator step for the neurons
for (unsigned int i = m_num_inputs; i < m_neurons.size(); i++)
{
double t_const = a_dtime / m_neurons[i].m_timeconst;
m_neurons[i].m_membrane_potential = (1.0 - t_const)
* m_neurons[i].m_membrane_potential
+ t_const * m_neurons[i].m_activesum;
}
// Now loop nodes_activesums, pass the signals through the activation function
// and store the result back to nodes_activations
// also skip inputs since they do not get an activation
for (unsigned int i = m_num_inputs; i < m_neurons.size(); i++)
{
double x = m_neurons[i].m_membrane_potential + m_neurons[i].m_bias;
m_neurons[i].m_activesum = 0;
// Apply the activation function
double y = 0.0;
switch (m_neurons[i].m_activation_function_type)
{
case SIGNED_SIGMOID:
y = af_sigmoid_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_SIGMOID:
y = af_sigmoid_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case TANH:
y = af_tanh(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case TANH_CUBIC:
y = af_tanh_cubic(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case SIGNED_STEP:
y = af_step_signed(x, m_neurons[i].m_b);
break;
case UNSIGNED_STEP:
y = af_step_unsigned(x, m_neurons[i].m_b);
break;
case SIGNED_GAUSS:
y = af_gauss_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_GAUSS:
y = af_gauss_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case ABS:
y = af_abs(x, m_neurons[i].m_b);
break;
case SIGNED_SINE:
y = af_sine_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_SINE:
y = af_sine_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case SIGNED_SQUARE:
y = af_square_signed(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case UNSIGNED_SQUARE:
y = af_square_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
case LINEAR:
y = af_linear(x, m_neurons[i].m_b);
break;
default:
y = af_sigmoid_unsigned(x, m_neurons[i].m_a, m_neurons[i].m_b);
break;
}
m_neurons[i].m_activation = y;
}
}