Per-Channel Softmax Map

Problem

Apply softmax over the class dimension of an H x W x C score map.

For each pixel subtract the maximum score before exponentiating for numerical stability, then normalize the exponentials and round each probability to 4 decimal places.

Hint

Subtracting the max keeps the exponentials from overflowing.

Test cases

case 1

in: [[[[1,1]]]]

expected: [[[0.5,0.5]]]

case 2

in: [[[[0,0,0]]]]

expected: [[[0.3333,0.3333,0.3333]]]

case 3

in: [[[[1000,1000]]]]

expected: [[[0.5,0.5]]]

case 4

in: [[[[1,3],[2,0]]]]

expected: [[[0.1192,0.8808],[0.8808,0.1192]]]

Starter codePython

def softmax_map(scores):
    # Your code here
    pass

4 test cases · run your code against them in the editor