SushiAI::NN::InitPolicy

InitFamily family = InitFamily::KAIMING

Which variance rule the weights follow.

Nonlinearity nonlinearity = Nonlinearity::RELU

The activation the hidden layers feed.

bool uniform = false

Draw uniformly rather than normally.

double negative_slope = 0.01

The leaky-ReLU slope, when that is the nonlinearity.

bool zero_final_layer = true

Zero the last weight and bias, so the initial loss is exactly ln(C).

std::uint64_t seed = 0x51757368'69414921ull

The Philox seed; the same seed gives the same weights.