3D Conv FLOPs

Problem

Count the multiply-accumulate floating point operations of a 3D convolution.

Each output element combines cin * kt * kh * kw inputs, and the factor 2 accounts for one multiply and one add. Return 2 * cin * cout * kt * kh * kw * t_out * h_out * w_out with bias excluded.

Hint

The kernel volume is kt * kh * kw; multiply by both channel counts and the output volume.

Test cases

case 1

in: [3,16,3,3,3,8,28,28]

expected: 16257024

case 2

in: [1,1,1,3,3,4,10,10]

expected: 7200

case 3

in: [64,64,3,3,3,2,14,14]

expected: 86704128

case 4

in: [2,4,2,2,2,3,5,5]

expected: 9600

Starter codePython

def conv3d_flops(cin, cout, kt, kh, kw, t_out, h_out, w_out):
    # Your code here
    pass

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