SushiAI::Train

enum class OptimizerPlacement

Whether the parameter update is part of the compiled step.

IN_PLAN

Recorded into the same plan as forward and backward; the default.

PER_STEP

Stepped on its own after every replay, with its own flush.

using SushiAI::Train::ObjectiveFactory = std::unique_ptr<IObjective> (*)(int64_t batch, int64_t classes, const Config::Value& attributes, std::string_view context)

Builds one objective from its configuration attributes.

constexpr std::array< ObjectiveEntry, 1 > OBJECTIVE_REGISTRY = { ObjectiveEntry{"cross_entropy", &make_cross_entropy}, }

Every objective a configuration file may name.

A file-scope constant table; nothing self-registers. See config/registry.hpp for why that is forced rather than preferred.

std::unique_ptr< IObjective > make_cross_entropy(int64_t batch, int64_t classes, const Config::Value &attributes, std::string_view context)

Builds a cross-entropy objective from attributes.

std::unique_ptr< IObjective > make_objective(std::string_view type, int64_t batch, int64_t classes, const Config::Value &attributes, std::string_view context)

Builds the objective named by type.

template <typename M>
TrainingGraph build_graph(M &module, const IObjective &objective, int64_t batch, int64_t features, DType dtype=DType::FLOAT32, Graph::FusionOptions fusion={}, const Graph::IPrecisionPolicy *precision=nullptr, bool loss_scaling=false)

Traces module into a full training graph and differentiates it.

template <typename M>
TrainingGraph build_classifier(M &module, int64_t batch, int64_t features, int64_t classes, DType dtype=DType::FLOAT32, Graph::FusionOptions fusion={}, const Graph::IPrecisionPolicy *precision=nullptr, bool loss_scaling=false)

Traces module into a cross-entropy classifier training graph.

std::vector< Optim::ParameterSlot > bind_parameters(const TrainingGraph &built, Graph::TensorPool &pool, Graph::PlanOptions options={})

Plans memory, allocates leaves, and pairs parameters with gradients.