torch:log-softmax

(torch:log-softmax a &key axis)

Differentiable log-softmax (linalg:log-softmax, the numerically stable half of a cross-entropy loss): computed as (x - max) - log(sum(exp(x - max))), never as log of torch:softmax. The backward pass is g - softmax(x) * sum(g).