torch:cross-entropy-loss

(torch:cross-entropy-loss logits targets &key ignore-index reduction)

Returns the cross entropy over raw logits as a scalar tensor (PyTorch's nn.CrossEntropyLoss). The logits have shape (... num-classes) -- the leading axes are flattened, so (batch seq vocab) works directly. It is computed from -log-softmax, which is the numerically stable form; do not pass softmax outputs.

The target is read one of two ways:

• class indices of the matching leading shape -- a number, a list, an index vector or a tensor. The loss is -log-softmax picked at the target class.
• class probabilities -- a tensor or array of the logits' own shape, PyTorch's soft-label form. The loss is -sum(target * log-softmax(logits)) per position, and the gradient flows into the target too when it requires one. A LIST is always class indices, so the probability spelling needs a tensor or an array.

:ignore-index k drops every position whose class-index target is k from both the sum and the mean's denominator, which is what keeps padding positions from contributing; like PyTorch it does not apply to probability targets. :reduction :sum adds instead of averaging and :reduction :none returns the per-position tensor.