IFusedReadoutLossLayer<T>

Interfaces · AiDotNet.Interfaces

A readout layer that can compute loss(Forward(input), target) as one fused, differentiable operation without materialising its output on the autodiff tape - for example a language-model head fused with cross-entropy (IEngine.FusedLinearCrossEntropyWithLogits), whose output is a [rows, vocabulary] logits tensor. Local-objective credit rules evaluate the readout once per hidden layer, so the decomposed path keeps one such tensor (and its softmax intermediates) per layer alive until the backward pass.

Methods

Method Summary
Tensor<T> TryComputeFusedLoss(Tensor<T> input, Tensor<T> target, ILossFunction<T> loss) Returns the scalar loss equal to loss.ComputeTapeLoss(Forward(input), target) reduced to a scalar, computed by a fused operation recorded on the current tape, or null when this loss / target combination is not supported (the caller then…