torch:multinomial

(torch:multinomial probs &key num-samples replacement)

num-samples indices drawn from each row of probs, whose LAST axis holds the weights (PyTorch's torch.multinomial). The weights need not sum to 1 -- each row is normalized -- and must be non-negative. A rank-1 input answers a (num-samples) index array, a rank-n input its own shape with the last axis replaced by num-samples. Non-differentiable, and a RAW linalg array rather than a tensor.

Without :replacement t an index already drawn cannot be drawn again within the same row, which is PyTorch's default; num-samples must then not exceed the number of weights. The draw comes from the SEEDED linalg:seed generator, so a sampling run reproduces on every backend.