# Set up evaluation
population_size = 20
epochs = 60
def const_epochs_60(individual, gen_index, prev_evaluations):
return epochs
# original paper
learning_rate_decay_points = [1, 149, 249]
relative_decay_points = [i / 350 for i in learning_rate_decay_points]
# parse that to out number of epochs
learning_rate_decay_points = [
int(np.ceil(i * epochs)) for i in relative_decay_points
]
data = cga.CIFAR10()
pickling_ga.run('',
data,
folder,
population_size=population_size,
search_space=search_space,
epoch_fn=const_epochs_60,
learning_rate_decay_points=learning_rate_decay_points,
fitness_punishment_per_hour=0.05)
start = 30
end = maxepochs
maxgens = 20
# y = ax + b
diff = end - start
b = start
a = diff / (maxgens - 1)
return round(a * gen_index + b)
# original paper
learning_rate_decay_points = [1, 149, 249]
relative_decay_points = [i / 350 for i in learning_rate_decay_points]
# parse that to out number of epochs
learning_rate_decay_points = [
int(np.ceil(i * maxepochs)) for i in relative_decay_points
]
data = cga.CIFAR10()
pickling_ga.run('',
data,
folder,
population_size=population_size,
search_space=search_space,
epoch_fn=linear_epochs_30_70,
learning_rate_decay_points=learning_rate_decay_points)
def const_epochs_60(individual, gen_index, prev_evaluations):
return epochs
def create_individual():
return cga.CNN.random_approach2(
input_shape=data.example.shape,
num_classes=data.num_classes,
search_space=search_space
)
# original paper
learning_rate_decay_points = [1, 149, 249]
relative_decay_points = [i/350 for i in learning_rate_decay_points]
# parse that to out number of epochs
learning_rate_decay_points = [int(np.ceil(i * epochs))
for i in relative_decay_points]
data = cga.CIFAR10()
pickling_ga.run('approach-2__', data, folder,
population_size=population_size,
search_space=search_space,
epoch_fn=const_epochs_60,
learning_rate_decay_points=learning_rate_decay_points,
create_individual=create_individual
)