N_SAMPLES = np.array([1e5, 1e6, 1e7], dtype=int)
N_FEATURES = np.array([1, 5, 10], dtype=int)
# Create several array for the data
if dataset == 'random':
array_data = [
misc.generate_samples(ns, nf, RND_SEED) for ns in N_SAMPLES
for nf in N_FEATURES
]
elif dataset == 'cover_type':
array_data = [misc.load_cover_type(RND_SEED)]
elif dataset == 'higgs':
# Select a subset of samples
array_data = [
misc.load_higgs(random_state=RND_SEED, n_samples=ns)
for ns in N_SAMPLES
]
else:
raise ValueError('The dataset is not known. The possible choices are:'
' random')
# Save only the time for the moment
res_lgbm = [[data[0].shape, p,
misc.bench(bench_lgbm, data, n=n_try, **p)]
for p in params_list for data in array_data]
# Check that the path is existing
if not os.path.exists(store_dir):
os.makedirs(store_dir)
from sklearn.model_selection import ParameterGrid
sys.path.insert(0, '../datasets')
from misc import load_higgs
configuration_path = "../params_benchmark/parameters_higgs.conf"
config_name = 'lightgbm'
with open(configuration_path, 'r') as stream:
params = yaml.load(stream)[config_name]
params = {
key: (value if isinstance(value, list) else [value])
for key, value in params.items()
}
params_grid = list(ParameterGrid(params))
params_selected = [d for d in params_grid if d['max_depth'] == 8][0]
N_SAMPLES = 1e7
data = load_higgs(random_state=42, n_samples=int(N_SAMPLES))
# Extract the parameter required for the dataset
max_bin = params_selected.pop('max_bin')
lgbm_training = lgb.Dataset(data[0], label=data[1], max_bin=max_bin)
n_est = params_selected.pop('n_estimators')
# Create the number of leafs depending of the max depth
params_selected['num_leaves'] = np.power(2, params_selected['max_depth'] - 1)
# Do not limit the depth of the trees
params_selected['max_depth'] = -1
gbm = lgb.train(params_selected, lgbm_training, num_boost_round=n_est)