def run_grid_search_multicore():
if not os.path.exists(os.path.join(data_dir(), dataset + '_train.libsvm')):
dataset_info = data_info(dataset)
get_file(dataset, origin=dataset_info['origin'], untar=True, md5_hash=dataset_info['md5_hash'])
params = {'regular_param': [3.8639876352395671e-06],
'gamma': [0.015625]}
candidate_params_lst = list(ParameterGrid(params))
pool = mp.Pool(processes=20) # maximum of workers
for candidate_params in candidate_params_lst:
pool.apply_async(run_one_candidate, args=(candidate_params,), callback=log_result)
# for candidate_params in candidate_params_lst:
# run_one_candidate(candidate_params)
pool.close()
pool.join()
if len(candidate_params_lst) > 1:
print("========== FINAL RESULT ==========")
idx_best = np.argmin(np.array(mistake_rate_lst))
print('Data set: {}'.format(dataset))
print('Oracle: {}'.format('budget'))
print('Loss func: {}'.format(loss_func))
print('Best mistake rate: {}'.format(mistake_rate_lst[idx_best]))
print('Best params: {}'.format(run_param_lst[idx_best]))
print('Time per candidate param: {}'.format(time_lst[idx_best]))
def load_mnist(shuffle_data=True, randseed='default'):
train_path = get_file("mnist_train.libsvm", origin=remote_data_dir() + "/mnist_train.libsvm",
cache_subdir="demo")
test_path = get_file("mnist_test.libsvm", origin=remote_data_dir() + "/mnist_test.libsvm",
cache_subdir="demo")
x_train, y_train = load_svmlight_file(train_path, n_features=784)
x_test, y_test = load_svmlight_file(test_path, n_features=784)
x_train = x_train.toarray() / 255.0
x_test = x_test.toarray() / 255.0
if shuffle_data:
shuffle(x_train, y_train, randseed=randseed)
return (x_train, y_train), (x_test, y_test)
def load_cifar10(shuffle_data=True, randseed='default'):
train_path = get_file("cifar10_5k_train.pkl",
origin=remote_data_dir() + "/cifar10_5k_train.pkl",
cache_subdir="demo")
test_path = get_file("cifar10_1k_test.pkl",
origin=remote_data_dir() + "/cifar10_1k_test.pkl",
cache_subdir="demo")
tmp = pickle.load(open(train_path, "rb"))
x_train, y_train = tmp['data'] / 255.0, tmp['labels']
tmp = pickle.load(open(test_path, "rb"))
x_test, y_test = tmp['data'] / 255.0, tmp['labels']
if shuffle_data:
shuffle(x_train, y_train, randseed=randseed)
return (x_train, y_train), (x_test, y_test)
def load_housing(shuffle_data=True, randseed='default'):
train_path = get_file("housing_scale_train.libsvm",
origin=remote_data_dir() + "/housing_scale_train.libsvm",
cache_subdir="demo")
test_path = get_file("housing_scale_test.libsvm",
origin=remote_data_dir() + "/housing_scale_test.libsvm",
cache_subdir="demo")
x_train, y_train = load_svmlight_file(train_path, n_features=13)
x_test, y_test = load_svmlight_file(test_path, n_features=13)
x_train = x_train.toarray()
x_test = x_test.toarray()
if shuffle_data:
shuffle(x_train, y_train, randseed=randseed)
return (x_train, y_train), (x_test, y_test)
def run_grid_search_multicore(
create_obj_func, params_gridsearch, attribute_names, dataset=None, num_workers=4, file_config=None,
num_runs=3, cross=0, num_features=None, full_dataset=None, keep_vars=[], ind_test=None, max_torrance=1):
if dataset is None:
if len(sys.argv) > 2:
dataset = sys.argv[2]
else:
raise Exception('Not specify dataset')
params_gridsearch = parse_arguments(params_gridsearch, True)
# print(params_gridsearch)
file_config, params_gridsearch = extract_param('file_config', file_config, params_gridsearch)
num_workers, params_gridsearch = extract_param('num_workers', num_workers, params_gridsearch)
num_runs, params_gridsearch = extract_param('num_runs', num_runs, params_gridsearch)
cross, params_gridsearch = extract_param('cross', cross, params_gridsearch)
num_features, params_gridsearch = extract_param('num_features', num_features, params_gridsearch)
full_dataset, params_gridsearch = extract_param('full_dataset', full_dataset, params_gridsearch)
ind_test, params_gridsearch = extract_param('ind_test', ind_test, params_gridsearch)
max_torrance, params_gridsearch = extract_param('max_torrance', max_torrance, params_gridsearch)
if ind_test is not None:
if full_dataset is None:
ind_test = dataset
else:
ind_test = full_dataset
if full_dataset is None:
full_dataset = dataset
if not os.path.exists(os.path.join(data_dir(), full_dataset + '_train.libsvm')):
dataset_info = data_info(full_dataset)
get_file(full_dataset, origin=dataset_info['origin'], untar=True, md5_hash=dataset_info['md5_hash'])
candidate_params_lst = list(ParameterGrid(params_gridsearch))
grid_search = True
if len(candidate_params_lst) == 1:
grid_search = False
pool = mp.Pool(num_workers) # maximum of workers
result_lst = []
for candidate_params in candidate_params_lst:
result = pool.apply_async(
run_one_candidate,
args=(
create_obj_func, candidate_params, dataset, attribute_names, file_config, num_runs, cross,
num_features, keep_vars, ind_test, grid_search, max_torrance, online),
callback=log_result
)
result_lst.append(result)
for result in result_lst:
result.get()
pool.close()
pool.join()
if len(candidate_params_lst) > 1:
print("========== FINAL RESULT ==========")
if online:
idx_best = np.argmin(np.array(mistake_rate_lst))
else:
idx_best = np.argmax(np.array(test_acc_lst))
print('Data set: {}'.format(dataset))
print('Best testid: {}'.format(testid_lst[idx_best]))
if online:
print('Best mistake rate: {}'.format(mistake_rate_lst[idx_best]))
else:
print('Best err on training set: {}'.format(1-train_acc_lst[idx_best]))
print('Best err on valid set: {}'.format(1-test_acc_lst[idx_best]))
print('Best params: {}'.format(run_param_lst[idx_best]))
if cross > 0:
print('Run the best one')
num_runs_for_best = num_runs
if num_runs < 3:
num_runs_for_best = 3
best_result = run_one_candidate(
create_obj_func, run_param_lst[idx_best], full_dataset, attribute_names, file_config, num_runs_for_best,
cross=0, num_features=num_features, keep_vars=keep_vars, online=online)
# best_result['gridsearch_time'] = np.sum(np.array(time_lst))
log_result(best_result)