MBConv Parameter Count

Problem

Count the weights of an MBConv block without biases: expansion 1x1 (in_channels * hidden), depthwise kernel (hidden * k * k) and projection 1x1 (hidden * out_channels), where hidden = in_channels * expand_ratio.

The signature is mbconv_param_count(in_channels, out_channels, kernel_size, expand_ratio). Return an integer.

Hint

Depthwise convolution uses one kernel per channel, so its weight count is small.

Test cases

case 1

in: [24,24,3,6]

expected: 8208

case 2

in: [16,24,3,6]

expected: 4704

case 3

in: [1,1,3,1]

expected: 11

case 4

in: [32,16,5,4]

expected: 9344

Starter codePython

def mbconv_param_count(in_channels, out_channels, kernel_size, expand_ratio):
    # Your code here
    pass

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