public class SparseCategoricalCrossentropy<T extends AbstractFloat = Float> implements ClassificationLoss<T>
Mean multiclass cross-entropy from logits or normalized probabilities and integer labels. Log-sum-exp is evaluated after subtracting each row maximum. No one-hot tensor or softmax Jacobian is materialized. Reduction is performed in FP64.
loss = SparseCategoricalCrossentropy<Float>().compute(logits, labels)
| Modifier and Type | Member | Description |
|---|---|---|
| public readonly property | fromLogitsfromLogits as Boolean |
Whether inputs are unnormalized logits rather than probabilities. |
| Modifier and Type | Member | Description |
|---|---|---|
| public | SparseCategoricalCrossentropySparseCategoricalCrossentropy(fromLogits as Boolean = true) |
Selects stable log-sum-exp mode or probability mode. |