class

`SushiAI::Train::CrossEntropyObjective`

Multi-class cross-entropy over one-hot targets, and accuracy.

Declared in: include/SushiAI/train/objective.hpp
Inherits: SushiAI::Train::IObjective

Emits nll_loss(log_softmax(output), targets) - the same two operations in the same order as the code this replaced, so the existing suite is the regression test.

Public member functions

CrossEntropyObjective(int64_t batch, int64_t classes, DType dtype=DType::FLOAT32) noexcept

Constructs the objective for a fixed batch and class count.

Parameters

batch: Samples per step.
classes: How many classes.
dtype: The element type the target leaf carries.

virtual void declare(const Graph::GraphBuilder &builder, ObjectiveLeaves &leaves) const override

Declares whatever leaves the graph needs beyond the model's input.

Parameters

builder: Where the leaves go.
leaves: Where to record them, by role.

virtual Graph::ValueId loss(const Graph::GraphBuilder &builder, const Graph::Graph &graph, Graph::ValueId output, const ObjectiveLeaves &leaves) const override

Computes scalar loss from model output.

Parameters

builder: Where the nodes go.
graph: The graph, for checking shapes.
output: The model output value.
leaves: Declared auxiliary leaves.

Returns

The scalar loss value identifier.

virtual std::optional< Metric > score(const Data::Batch &batch, const SushiBLAS::Tensor &output) const override

An optional per-step metric: accuracy, mAP, an L2 residual, or none.

Parameters

batch: The batch that was just stepped.
output: The model's output tensor.

Returns

The metric, or std::nullopt when this objective offers none.

virtual std::string_view name() const noexcept override

A short name for a run report.