pop_size = params['pop_size']

max_length = params['max_length']

max_stride2 = params['max_stride2']

image_channel = params['image_channel']

max_output_channel = params['max_output_channel']

population = []

for _ in range(pop_size):

num_net = int(max_length)

part2 = []

p_valid = np.random.random()

for i in range(num_net):

p_ = np.random.random()

if p_ <= p_valid:

num_feature_maps = np.random.randint(0, max_output_channel)

else:

p1 = np.random.random()

if p1 <= 0.5:

num_feature_maps = np.random.randint(-30, 0)

else:

num_feature_maps = np.random.randint(max_output_channel, max_output_channel+30)

part2.append(num_feature_maps)

num_stride2 = np.random.randint(0, max_stride2 + 1)

num_stride1 = num_net - num_stride2

# find the position where the pooling layer can be connected

availabel_positions = list(idx for idx in range(0, num_net) if 0<=part2[idx]<max_output_channel) #only consider those valid layers as possible strided layers

if len(availabel_positions) == 0:

part2[0] = np.random.randint(0, max_output_channel)

availabel_positions = [0]

np.random.shuffle(availabel_positions)

np.random.shuffle(availabel_positions)

while len(availabel_positions) < num_stride2:

supp_list = [idx for idx in range(0, num_net) if idx not in availabel_positions]

availabel_positions.append(np.random.choice(supp_list, 1)[0])

select_positions = np.sort(availabel_positions[0:num_stride2]) # the positions of pooling layers in the net

part1 = []

for i in range(num_net):

if i in select_positions:

code_stride2 = np.random.randint(3, 6)

part1.append(code_stride2)

else:

code_stride1 = np.random.randint(0, 3)

part1.append(code_stride1)

population.append([part1, part2])

params = {}

params['pop_size'] = 20

params['max_length'] = 30

params['max_stride2'] = 4

params['image_channel'] = 1

params['max_output_channel'] = 256

pop = initialize_population(params)