Return the scalar mean categorical cross-entropy:

• Z holds logits of shape (n, k)
• y holds integer labels of shape (n,)
• compute it via log-sum-exp, so it stays finite when a correct-class probability underflows to 0

Input

Z =
[[1. 2. 3.]
 [3. 1. 0.]]
y = [2 0]

Output

0.2887259920003331