Softmax over logits

A model's raw scores become probabilities through softmax. Compute it, safely.

What to build

A program that turns logits into probabilities with a numerically stable softmax.

Requirements

  • Print the probabilities in the same order, each with 3 decimal places.
  • It must work for very large logits, such as 1000.

Input

One line of logits.

Output

One line.

Constraints

1 to 1000 logits, each between −10⁴ and 10⁴.

Example

Input:

0 0

Output:

0.500 0.500

Why it works

Subtracting the maximum keeps every exponent at most 0, so e^z never overflows, and the shift cancels out when dividing by the sum.

Examples

Example 1 · Input
0 0
Expected output
0.500 0.500
Example 2 · Input
1 2 3
Expected output
0.090 0.245 0.665

Prompts inside input() are ignored when checking, so input("n: ") is fine.

Hints

Hint 1

Subtract the largest logit from every logit first.

Hint 2

Exponentiate each, then divide by their sum.

Solution

Show one correct solution (try it yourself first)
import math

logits = list(map(float, input().split()))
top = max(logits)
exps = [math.exp(z - top) for z in logits]
total = sum(exps)
print(*[f"{e / total:.3f}" for e in exps])

Lesson: Next-token prediction →