Function nll_loss 

pub fn nll_loss<B>(
    log_probs: &Tensor<B>,
    targets: &Tensor<B>,
) -> Result<Tensor<B>, Error>
where
    B: Backend,

Negative Log-Likelihood Loss with integer class indices.

Computes: -mean(log_probs[i, target[i]]) for each sample i.

Arguments

• log_probs: log-probabilities [batch, num_classes] (output of log_softmax)
• targets: class indices as f64 [batch] — each value in 0..num_classes

Typically used as: nll_loss(&logits.log_softmax(1)?, &targets)

Note: unlike cross_entropy_loss which takes one-hot targets, this takes integer class indices (more memory efficient).