EfficientNet Compound Scaling

Problem

Scale width, depth and resolution together with the compound coefficients: [base_width * alpha^phi, base_depth * beta^phi, base_res * gamma^phi].

The signature is efficientnet_compound_scaling(base_width, base_depth, base_res, phi, alpha=1.2, beta=1.1, gamma=1.15).

Hint

Compound scaling beat scaling just one dimension because accuracy depends on all three jointly.

case 1

in: [1,1,224,1]

expected: [1.2,1.1,257.6]

case 2

in: [1,1,224,2]

expected: [1.44,1.21,296.24]

case 3

in: [1,1,224,0]

expected: [1,1,224]

case 4

in: [1.5,2,128,1]

expected: [1.8,2.2,147.2]

Starter codePython

def efficientnet_compound_scaling(base_width, base_depth, base_res, phi, alpha=1.2, beta=1.1, gamma=1.15):
    # Your code here
    pass

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