def NAS_EA_FA(dataset: str):
cur_time_budget = 0
history2acc = {}
best_valid_acc, best_test_acc, times = [0.0], [0.0], [0.0]
x_train, y_train = [], []
total_population = [Individual() for _ in range(POPULATION_SIZE)]
while cur_time_budget <= MAX_TIME_BUDGET:
population = sorted(total_population, key=lambda x: x.fitness, reverse=True)
new_population = []
num = 0
for individual in population:
valid_acc, test_acc, time = get_individual_data_from_history(individual, history2acc, dataset)
individual.fitness = valid_acc
if time == 0.0:
continue
new_population.append(individual)
x_train.append(get_individual_sequences(individual))
y_train.append(valid_acc)
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
num += 1
if cur_time_budget > MAX_TIME_BUDGET or num >= 5:
break
if len(new_population) != 0:
population = new_population
gbm = XGBRegressor(eta=0.1)
gbm.fit(np.array(x_train), np.array(y_train))
total_population = []
for epoch in range(10):
new_population = []
for i in range(POPULATION_SIZE):
individual = cp.deepcopy(tournament_selection(population))
bitwise_mutation(individual)
individual.fitness = gbm.predict(np.array([get_individual_sequences(individual)]))[0]
new_population.append(individual)
total_population.append(individual)
population = new_population
return best_valid_acc, best_test_acc, times
def random_search(dataset: str):
cur_time_budget = 0
best_valid_acc, best_test_acc, times = [0.0], [0.0], [0.0]
while cur_time_budget <= MAX_TIME_BUDGET:
individual = Individual()
valid_acc, test_acc, time = get_individual_data(individual, dataset)
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
return best_valid_acc, best_test_acc, times
def regularized_evolution_algorithm(dataset: str):
cur_time_budget = 0
history2acc = {}
best_valid_acc, best_test_acc, times = [0.0], [0.0], [0.0]
population = [Individual() for _ in range(POPULATION_SIZE)]
for individual in population:
valid_acc, test_acc, time = get_individual_data_from_history(
individual, history2acc, dataset)
individual.fitness = valid_acc
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
while cur_time_budget <= MAX_TIME_BUDGET:
individual = cp.deepcopy(tournament_selection(population))
bitwise_mutation(individual)
valid_acc, test_acc, time = get_individual_data_from_history(
individual, history2acc, dataset)
individual.fitness = valid_acc
population.append(individual)
population.pop(0)
if time == 0.0:
continue
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
if cur_time_budget > MAX_TIME_BUDGET:
break
return best_valid_acc, best_test_acc, times
def random_search(dataset: str):
cur_time_budget = 0
best_valid_acc, best_test_acc, times = [0.0], [0.0], [0.0]
history_hash = set()
while cur_time_budget <= MAX_TIME_BUDGET:
individual = Individual()
individual_hash = get_individual_hash(individual)
if individual_hash in history_hash:
continue
else:
history_hash.add(individual_hash)
valid_acc, test_acc, time = get_individual_data(individual, dataset)
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
return best_valid_acc, best_test_acc, times
def sample_models(n=100):
models = [Individual() for _ in range(n)]
return models
def NAS_EA_FA_V2(dataset: str):
cur_time_budget = 0
history2acc = {}
best_valid_acc, best_test_acc, times = [0.0], [0.0], [0.0]
x_train, y_train = [], []
total_population = [Individual() for _ in range(POPULATION_SIZE)]
while cur_time_budget <= MAX_TIME_BUDGET:
population = sorted(total_population, key=lambda x: x.fitness, reverse=True)
new_population = []
num = start_index = 0
for index in range(len(population)):
individual = population[index]
valid_acc, test_acc, time = get_individual_data_from_history(individual, history2acc, dataset)
individual.fitness = valid_acc
if time == 0.0:
continue
new_population.append(individual)
# x_train.append(get_individual_sequences(individual))
temp_sequences = get_all_isomorphic_sequences(individual)
x_train.extend(temp_sequences)
y_train.extend([valid_acc for _ in range(len(temp_sequences))])
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
num += 1
if cur_time_budget > MAX_TIME_BUDGET or num >= 3:
start_index = index
break
# diversity individual
max_dis_list = [get_min_distance(x_train, get_individual_sequences(indiv)) for indiv in
population[start_index + 1:]]
while num < 5 and cur_time_budget <= MAX_TIME_BUDGET:
max_dis, temp_index = 0, 0
for index in range(len(max_dis_list)):
if max_dis_list[index] > max_dis:
max_dis = max_dis_list[index]
temp_index = index
if max_dis == 0:
break
individual = population[temp_index + start_index + 1]
max_dis_list[temp_index] = 0
valid_acc, test_acc, time = get_individual_data_from_history(individual, history2acc, dataset)
individual.fitness = valid_acc
if time == 0.0:
continue
new_population.append(individual)
temp_sequences = get_all_isomorphic_sequences(individual)
x_train.extend(temp_sequences)
y_train.extend([valid_acc for _ in range(len(temp_sequences))])
if valid_acc > best_valid_acc[-1]:
best_valid_acc.append(valid_acc)
best_test_acc.append(test_acc)
else:
best_valid_acc.append(best_valid_acc[-1])
best_test_acc.append(best_test_acc[-1])
times.append(time)
cur_time_budget += time
num += 1
if len(new_population) != 0:
population = new_population
gbm = XGBRegressor(eta=0.1)
gbm.fit(np.array(x_train), np.array(y_train))
total_population = []
for epoch in range(10):
new_population = []
for i in range(POPULATION_SIZE):
individual = cp.deepcopy(tournament_selection(population))
bitwise_mutation(individual)
individual.fitness = gbm.predict(np.array([get_individual_sequences(individual)]))[0]
new_population.append(individual)
total_population.append(individual)
population = new_population
return best_valid_acc, best_test_acc, times