Contents

struct

SushiAI::Optim::SGDOptions

SGD's hyperparameters, with PyTorch's defaults where they have one.

Declared in
include/SushiAI/optim/sgd.hpp

Public attributes

double learning_rate = 1e-2

Step size.

double momentum = 0.0

Momentum coefficient; 0 disables the buffer entirely.

double dampening = 0.0

Damping applied to the gradient from step 2 onward.

double weight_decay = 0.0

Coupled L2 penalty, added to the gradient.

bool nesterov = false

Use Nesterov's look-ahead; requires momentum > 0 and dampening == 0.

UpdateKernel kernel = UpdateKernel::FUSED

One fused kernel per parameter, or the separate ops.

FUSED is the default and the shipping path; SEPARATE exists to be compared against. See optimizer.hpp.