Thin wrapper to give module-like ergonomics when wiring a loss into a pipeline. The single-argument forward is intentionally unsupported; call forward(preds, targets, ctx).

MAELoss

class MAELoss : Loss

Mean Absolute Error (L1) Loss.

MSELoss

class MSELoss : Loss

PoissonLoss

class PoissonLoss @JvmOverloads constructor(logInput: Boolean = true, epsilon: Float = 1.0E-8f) : Loss

Poisson Negative Log Likelihood Loss.

Reduction

enum Reduction : Enum<Reduction>

SparseCategoricalCrossEntropyLoss

class SparseCategoricalCrossEntropyLoss @JvmOverloads constructor(dim: Int = -1) : Loss

Sparse Categorical Cross-Entropy Loss.

SquaredHingeLoss

class SquaredHingeLoss @JvmOverloads constructor(margin: Float = 1.0f) : Loss

Squared Hinge Loss.

Functions

evaluateLoss

fun <T : DType, V> evaluateLoss(model: Module<T, V>, loss: Loss, inputs: Tensor<T, V>, targets: Tensor<out DType, *>, ctx: ExecutionContext, reduction: Reduction = Reduction.MEAN): Tensor<T, V>

Utility helpers to evaluate a loss given a model, inputs, and targets. Keeps loss evaluation explicit and separate from the network-building DSL.