Negative Log-Likelihood

Negative log-likelihood for one-hot targets when the input is already log-probabilities.

noncomputable def Proofs.Autograd.TapeNodes.nllOneHotLast {Γ : List Spec.Shape} {m n : ℕ} (logProbs target : Idx Γ (Spec.Shape.dim m (Spec.Shape.dim n Spec.Shape.scalar))) : Node Γ Spec.Shape.scalar

Negative log-likelihood loss for log-probabilities and one-hot targets of shape (m×n).

Forward: -(1/m) * ⟪target, logProbs⟫

This is the natural primitive loss that cross_entropy reduces to after log_softmax.

noncomputable def Proofs.Autograd.TapeNodes.nllOneHotLastFderiv {Γ : List Spec.Shape} {m n : ℕ} (logProbs target : Idx Γ (Spec.Shape.dim m (Spec.Shape.dim n Spec.Shape.scalar))) : NodeFDerivCorrect (nllOneHotLast logProbs target)

NodeFDerivCorrect for nll_one_hot_last (negative log-likelihood with one-hot targets).