NSGA-Net: Neural Architecture Search using Multi-Objective Genetic Algorithm

Generate random architectures

from search.micro_encoding import decode
import numpy.random as rd

ops=rd.randint(0,8,16)
links0=rd.choice([0,1], size=2, replace=False)
links1=rd.choice([0,1,2], size=2, replace=False)
links2=rd.choice([0,1,2,3], size=2, replace=False)
links3=rd.choice([0,1,2,3,4], size=2, replace=False)
links0r=rd.choice([0,1], size=2, replace=False)
links1r=rd.choice([0,1,2], size=2, replace=False)
links2r=rd.choice([0,1,2,3], size=2, replace=False)
links3r=rd.choice([0,1,2,3,4], size=2, replace=False)
genome = [[[[ops[0], links0[0]], [ops[1], links0[1]]], [[ops[2], links1[0]], [ops[3], links1[1]]], [[ops[4], links2[0]], [ops[5], links2[1]]], [[ops[6], links3[0]], [ops[7], links3[1]]]],
         [[[ops[8], links0r[0]], [ops[9], links0r[1]]], [[ops[10], links1r[0]], [ops[11], links1r[1]]], [[ops[12], links2r[0]], [ops[13], links2r[1]]], [[ops[14], links3r[0]], [ops[15], links3r[1]]]]]
genotype = decode(genome)

Search

python search/evolution_search.py 
--init_channels 16 
--layers 8 
--epochs 20 
--n_offspring 20 
--n_gens 30 
--search_space micro 
--save test 
--data_path /data # path to data
--dataset CIFAR10 # choose between CIFAR10, CIFAR100, Sport8, MIT67 and flowers102

Augment

python validation/train.py 
--dataset CIFAR10 # choose between CIFAR10, CIFAR100, Sport8, MIT67 and flowers102
--net_type micro 
--layers 20 # 20 for CIFAR10 and CIFAR100, 8 for Sport8, MIT67 and flowers102
--init_channels 34 
--filter_increment 4  
--cutout 
--auxiliary 
--batch_size 96 
--droprate 0.2 
--SE 
--epochs 600 
--genotype genotype 
--data datapath 
--save test 
--path test