def benchmark_cost_function(data_sizes):
polygon = pickle.load(open('utils/good_path_for_rectangle.pickle', 'rb'))
save_folder = generate_folder_name()
iterations = 10
results = np.empty((len(data_sizes), iterations, 2), dtype=float)
for i, dsize in enumerate(data_sizes):
data = dg.generate_rectangle_set(dsize)
print("\nRun {} with value {}".format(i+1, dsize))
# Compile functions and warm up GPU
acoc.cost_function_gpu(data.T, polygon)
for j in range(iterations):
utils.print_on_current_line('Iteration {}/{}'.format(j, iterations))
start_cpu = time.clock()
acoc.cost_function(data.T, polygon)
end_cpu = time.clock()
results[i][j][0] = end_cpu - start_cpu
start_gpu = time.clock()
acoc.cost_function_gpu(data.T, polygon)
end_gpu = time.clock()
results[i][j][2] = end_gpu - start_gpu
mean_results = np.mean(results, axis=1).T
acoc_plotter.plot_bar_chart_gpu_benchmark(mean_results, data_sizes, ['CPython', 'GPU'], save_folder, 'results')
np.set_printoptions(precision=7, suppress=False)
print("\nResults: \n{}".format(mean_results))
utils.save_object(mean_results, save_folder, 'results')
def _construct_polygon(self, data, plane_string, start_time, print_string):
cuda.to_device(data)
ant_scores = []
current_best_ant = []
last_level_up_or_best_ant = 0
matrix = AcocMatrix(data, tau_initial=self.config.tau_init)
current_best_score = 0
while matrix.level <= self.config.max_level:
start_vertex = get_random_weighted(matrix.edges)
if self.config.multi_level:
if (len(ant_scores) - last_level_up_or_best_ant
) > self.config.level_convergence_rate:
matrix.level_up(current_best_ant)
last_level_up_or_best_ant = len(ant_scores)
_ant = Ant(start_vertex)
_ant.move_ant()
while not _ant.at_target and not _ant.is_stuck:
_ant.move_ant()
if _ant.at_target:
ant_score = self._score(_ant.edges_travelled, data)
if ant_score > current_best_score:
current_best_ant = _ant.edges_travelled
current_best_score = ant_score
last_level_up_or_best_ant = len(ant_scores)
if self.config.plot:
plotter.plot_path_with_data(
current_best_ant,
data,
matrix,
save=True,
save_folder=osp.join(
self.save_folder,
'best_paths/{}/'.format(plane_string)),
file_name='ant' + str(len(ant_scores)))
self._put_pheromones(current_best_ant, current_best_score)
self._reset_at_random(matrix)
ant_scores.append(ant_score)
t_elapsed = utils.seconds_to_hms(time.time() - start_time)
utils.print_on_current_line(
"Level {}/{}, {}, Time elapsed: {}".format(
matrix.level, self.config.max_level, print_string,
t_elapsed))
return current_best_ant
def run(*args):
config = dict(CONFIG)
for conf in args:
config[conf[0]] = conf[1]
data = pickle.load(open('utils/data_sets.pickle', 'rb'))[config['data_set']]
number_runs = config['number_runs']
clf = acoc.PolyACO(config)
run_times = np.zeros(number_runs, dtype=float)
for i in range(number_runs):
iter_string = "Iteration: {}/{}".format(i + 1, number_runs)
start = time.clock()
clf.train(data, print_string=', ' + iter_string)
end = time.clock()
run_times[i] = end - start
utils.print_on_current_line(iter_string)
return np.mean(run_times)
def run(*args):
config = dict(CONFIG)
for conf in args:
config[conf[0]] = conf[1]
data = pickle.load(open('utils/data_sets.pickle',
'rb'))[config['data_set']]
number_runs = config['number_runs']
clf = acoc.PolyACO(config)
run_times = np.zeros(number_runs, dtype=float)
for i in range(number_runs):
iter_string = "Iteration: {}/{}".format(i + 1, number_runs)
start = time.clock()
clf.train(data, print_string=', ' + iter_string)
end = time.clock()
run_times[i] = end - start
utils.print_on_current_line(iter_string)
return np.mean(run_times)
def parameter_tester(parameter_name, values, save_folder=None):
if save_folder is None:
save_folder = utils.generate_folder_name()
print("\n\nExperiment for parameter '{}' with values {}".format(
parameter_name, values))
plt.clf()
all_scores = []
for index, v in enumerate(values):
print("Run {} with value {}".format(index + 1, v))
scores = run((parameter_name, v))
all_scores.append(scores)
utils.print_on_current_line('')
header = ','.join(str(s) for s in values)
result_str = header + '\n' + ','.join(["{:.4f}".format(sum(s) / CLASSIFIER_CONFIG.runs) for s in all_scores]) + \
'\n\n' + 'all scores:\n'
for a in all_scores:
result_str += ','.join('{:.4f}'.format(s) for s in a) + '\n'
utils.save_string_to_file(result_str,
parent_folder=save_folder,
file_name='result_' + parameter_name + '.txt')
def _construct_polygon(self, data, plane_string, start_time, print_string):
cuda.to_device(data)
ant_scores = []
current_best_ant = []
last_level_up_or_best_ant = 0
matrix = AcocMatrix(data, tau_initial=self.config.tau_init)
current_best_score = 0
while matrix.level <= self.config.max_level:
start_vertex = get_random_weighted(matrix.edges)
if self.config.multi_level:
if (len(ant_scores) - last_level_up_or_best_ant) > self.config.level_convergence_rate:
matrix.level_up(current_best_ant)
last_level_up_or_best_ant = len(ant_scores)
_ant = Ant(start_vertex)
_ant.move_ant()
while not _ant.at_target and not _ant.is_stuck:
_ant.move_ant()
if _ant.at_target:
ant_score = self._score(_ant.edges_travelled, data)
if ant_score > current_best_score:
current_best_ant = _ant.edges_travelled
current_best_score = ant_score
last_level_up_or_best_ant = len(ant_scores)
if self.config.plot:
plotter.plot_path_with_data(current_best_ant, data, matrix, save=True,
save_folder=osp.join(self.save_folder,
'best_paths/{}/'.format(plane_string)),
file_name='ant' + str(len(ant_scores)))
self._put_pheromones(current_best_ant, current_best_score)
self._reset_at_random(matrix)
ant_scores.append(ant_score)
t_elapsed = utils.seconds_to_hms(time.time() - start_time)
utils.print_on_current_line("Level {}/{}, {}, Time elapsed: {}".format(matrix.level, self.config.max_level, print_string, t_elapsed))
return current_best_ant
def benchmark_cost_function(data_sizes):
polygon = pickle.load(open('utils/good_path_for_rectangle.pickle', 'rb'))
save_folder = generate_folder_name()
iterations = 10
results = np.empty((len(data_sizes), iterations, 2), dtype=float)
for i, dsize in enumerate(data_sizes):
data = dg.generate_rectangle_set(dsize)
print("\nRun {} with value {}".format(i + 1, dsize))
# Compile functions and warm up GPU
acoc.cost_function_gpu(data.T, polygon)
for j in range(iterations):
utils.print_on_current_line('Iteration {}/{}'.format(
j, iterations))
start_cpu = time.clock()
acoc.cost_function(data.T, polygon)
end_cpu = time.clock()
results[i][j][0] = end_cpu - start_cpu
start_gpu = time.clock()
acoc.cost_function_gpu(data.T, polygon)
end_gpu = time.clock()
results[i][j][2] = end_gpu - start_gpu
mean_results = np.mean(results, axis=1).T
acoc_plotter.plot_bar_chart_gpu_benchmark(mean_results, data_sizes,
['CPython', 'GPU'], save_folder,
'results')
np.set_printoptions(precision=7, suppress=False)
print("\nResults: \n{}".format(mean_results))
utils.save_object(mean_results, save_folder, 'results')