def get_feature_vector(dataset, dataset_id, data_dir='./', task_type=None):
feature_dict = calculate_metafeatures(dataset,
dataset_id,
data_dir,
task_type=task_type)
sorted_keys = sorted(feature_dict.keys())
return [feature_dict[key] for key in sorted_keys]
def prepare_meta_dataset(meta_dir,
metric,
total_resource,
rep,
buildin_datasets,
buildin_algorithms,
task_type=None):
X, Y = list(), list()
sorted_keys = None
include_datasets = list()
for _dataset in buildin_datasets:
print('Processing', _dataset)
# Calculate metafeature for datasets.
try:
feature_dict = calculate_metafeatures(_dataset,
task_type=task_type)
except Exception as e:
print(e)
continue
if sorted_keys is None:
sorted_keys = sorted(feature_dict.keys())
meta_instance = [feature_dict[key] for key in sorted_keys]
X.append(meta_instance)
# Load partial relationship between algorithms.
scores = fetch_algorithm_runs(meta_dir, _dataset, metric,
total_resource, rep, buildin_algorithms)
Y.append(scores)
include_datasets.append(_dataset)
return X, Y, include_datasets
def fetch_algorithm_set(self, dataset, datanode=None):
input_vector = get_feature_vector(dataset, task_type=self.task_type)
if input_vector is None:
input_dict = calculate_metafeatures(dataset=datanode,
task_type=self.task_type)
sorted_keys = sorted(input_dict.keys())
input_vector = [input_dict[key] for key in sorted_keys]
preds = self.predict(input_vector)
idxs = np.argsort(-preds)
return [self.algorithms[idx] for idx in idxs]
def get_feature_vector(dataset, dataset_id, data_dir='./', task_type=None):
meta_dir = os.path.dirname(__file__) + '/../meta_resource/'
dataset_meta_feat_filename = meta_dir + 'meta_feature_dataset_%s.pkl' % dataset
if os.path.exists(dataset_meta_feat_filename):
with open(dataset_meta_feat_filename, 'rb') as f:
feature_vec = pickle.load(f)
return feature_vec
else:
feature_dict = calculate_metafeatures(dataset,
dataset_id,
data_dir,
task_type=task_type)
sorted_keys = sorted(feature_dict.keys())
return [feature_dict[key] for key in sorted_keys]
def prepare_meta_dataset(meta_dir,
metric,
total_resource,
rep,
buildin_datasets,
buildin_algorithms,
task_type=None):
X, Y = list(), list()
sorted_keys = None
include_datasets = list()
for _dataset in buildin_datasets:
print('Processing', _dataset)
dataset_meta_feat_filename = meta_dir + 'meta_feature_dataset_%s.pkl' % _dataset
if os.path.exists(dataset_meta_feat_filename):
with open(dataset_meta_feat_filename, 'rb') as f:
meta_instance = pickle.load(f)
else:
# Calculate metafeature for datasets.
try:
feature_dict = calculate_metafeatures(_dataset,
task_type=task_type)
except Exception as e:
print(e)
continue
if sorted_keys is None:
sorted_keys = sorted(feature_dict.keys())
meta_instance = [feature_dict[key] for key in sorted_keys]
with open(dataset_meta_feat_filename, 'wb') as f:
pickle.dump(meta_instance, f)
X.append(meta_instance)
# Load partial relationship between algorithms.
scores = fetch_algorithm_runs(meta_dir, _dataset, metric,
total_resource, rep, buildin_algorithms)
Y.append(scores)
include_datasets.append(_dataset)
return X, Y, include_datasets
'analcatdata_supreme', 'splice', 'abalone', 'spambase',
'winequality_white', 'waveform-5000(1)', 'waveform-5000(2)',
'page-blocks(1)', 'page-blocks(2)', 'cpu_act', 'optdigits', 'satimage',
'wind', 'musk', 'delta_ailerons', 'bank32nh', 'mushroom', 'puma8NH'
]
task_ids = list()
X = list()
result = dict()
sorted_keys = None
for _dataset in datasets:
print('Creating embedding for dataset - %s.' % _dataset)
# Calculate metafeature for datasets.
try:
feature_dict = calculate_metafeatures(_dataset,
task_type=MULTICLASS_CLS)
except Exception as e:
print(e)
continue
sorted_keys = sorted(feature_dict.keys())
meta_instance = [feature_dict[key] for key in sorted_keys]
X.append(meta_instance)
task_ids.append(_dataset)
result['task_ids'] = task_ids
result['dataset_embedding'] = np.array(X)
result['meta_feature_names'] = sorted_keys
with open('dataset_metafeatures.pkl', 'wb') as f:
pk.dump(result, f)