def evaluate():
iris = load_iris()
X, y = iris.data, iris.target
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=1)
try:
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
test_data = dm.get_data_node(X_test, y_test)
clf = Classifier(dataset_name='iris',
time_limit=150,
output_dir=save_dir,
ensemble_method=ensemble_method,
evaluation=eval_type,
metric='acc')
clf.fit(train_data)
clf.refit()
pred = clf.predict(test_data)
print('final score', clf.score(test_data))
except Exception as e:
return False
return True
def main():
time_limit = 60
print('==> Start to evaluate with Budget %d' % time_limit)
iris = load_iris()
X, y = iris.data, iris.target
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=1,
stratify=y)
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
test_data = dm.get_data_node(X_test, y_test)
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
add_classifier(UserDefinedDecisionTree)
clf = Classifier(time_limit=time_limit,
output_dir=save_dir,
enable_meta_algorithm_selection=False,
include_algorithms=['UserDefinedDecisionTree'],
ensemble_method=None,
metric='acc')
_start_time = time.time()
clf.fit(train_data)
print(clf.summary())
pred = clf.predict(test_data)
print(accuracy_score(test_data.data[1], pred))
shutil.rmtree(save_dir)
def fit(self, X, y):
if isinstance(X, pd.DataFrame):
self.headers = X.columns
X = np.array(X)
y = np.array(y)
self.origin_X = X
self.origin_y = y
if (np.sum(np.isnan(X)) == 0):
self.impute = False
self.impute_method = ['mean']
self.preprocess(X, y)
for key in self.impute_operator:
X_train = self.train_data[key]
self.dm = DataManager(X_train, y)
train_data = self.dm.get_data_node(X_train, y)
self.mdl[key] = soln_Regressor(
time_limit=self.time_limit / len(self.impute_method),
output_dir=self.output_dir,
ensemble_method=self.ensemble_method,
evaluation=self.evaluation,
metric=self.metric,
n_jobs=self.n_jobs)
self.mdl[key].fit(train_data)
return 0
def test_rgs():
time_limit = 60
print('==> Start to evaluate with Budget %d' % time_limit)
ensemble_method = 'bagging'
eval_type = 'holdout'
boston = load_boston()
X, y = boston.data, boston.target
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=1)
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
test_data = dm.get_data_node(X_test, y_test)
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
rgs = Regressor(metric='mse',
ensemble_method=ensemble_method,
enable_meta_algorithm_selection=False,
evaluation=eval_type,
time_limit=time_limit,
output_dir=save_dir)
rgs.fit(train_data)
print(rgs.summary())
pred = rgs.predict(test_data)
print(mean_squared_error(test_data.data[1], pred))
shutil.rmtree(save_dir)
def main():
ensemble_method = None
time_limit = 120
print('==> Start to evaluate with Budget %d' % time_limit)
eval_type = 'holdout'
boston = load_boston()
X, y = boston.data, boston.target
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=1)
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
test_data = dm.get_data_node(X_test, y_test)
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
rgs = Regressor(metric='mse',
ensemble_method=ensemble_method,
evaluation=eval_type,
time_limit=time_limit,
output_dir=save_dir)
rgs.fit(train_data)
pred = rgs.predict(test_data)
print(mean_squared_error(test_data.data[1], pred))
def load_data(dataset,
data_dir='./',
datanode_returned=False,
preprocess=True,
task_type=None):
dm = DataManager()
if task_type is None:
data_path = data_dir + 'data/datasets/%s.csv' % dataset
elif task_type in CLS_TASKS:
data_path = data_dir + 'data/cls_datasets/%s.csv' % dataset
elif task_type in REG_TASKS:
data_path = data_dir + 'data/rgs_datasets/%s.csv' % dataset
else:
raise ValueError("Unknown task type %s" % str(task_type))
# if dataset in ['credit_default']:
# data_path = data_dir + 'data/datasets/%s.xls' % dataset
# Load train data.
if dataset in [
'higgs', 'amazon_employee', 'spectf', 'usps', 'vehicle_sensIT',
'codrna'
]:
label_column = 0
elif dataset in ['rmftsa_sleepdata(1)']:
label_column = 1
else:
label_column = -1
if dataset in ['spambase', 'messidor_features']:
header = None
else:
header = 'infer'
if dataset in ['winequality_white', 'winequality_red']:
sep = ';'
else:
sep = ','
train_data_node = dm.load_train_csv(
data_path,
label_col=label_column,
header=header,
sep=sep,
na_values=["n/a", "na", "--", "-", "?"])
if preprocess:
pipeline = FEPipeline(fe_enabled=False,
metric='acc',
task_type=task_type)
train_data = pipeline.fit_transform(train_data_node)
else:
train_data = train_data_node
if datanode_returned:
return train_data
else:
X, y = train_data.data
feature_types = train_data.feature_types
return X, y, feature_types
def test_cls():
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
time_limit = 60
print('==> Start to evaluate with Budget %d' % time_limit)
ensemble_method = 'bagging'
eval_type = 'holdout'
iris = load_iris()
X, y = iris.data, iris.target
X_train, X_test, y_train, y_test = train_test_split(X,
y,
test_size=0.33,
random_state=1,
stratify=y)
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
test_data = dm.get_data_node(X_test, y_test)
clf = Classifier(time_limit=time_limit,
output_dir=save_dir,
enable_meta_algorithm_selection=False,
ensemble_method=ensemble_method,
ensemble_size=10,
evaluation=eval_type,
metric='acc')
clf.fit(train_data)
print(clf.summary())
pred = clf.predict(test_data)
print(accuracy_score(test_data.data[1], pred))
shutil.rmtree(save_dir)
def operate(self, data):
dm = DataManager(data)
data = dm.get_data_node(data, [])
for fe in self.list:
if fe is not None:
data = fe.operate(data)
return data.data[0]
def evaluate_fe_pipeline():
from solnml.utils.data_manager import DataManager
dm = DataManager()
# file_path = "data/proprocess_data.csv"
file_path = 'data/a9a/dataset_183_adult.csv'
dm.load_train_csv(file_path)
pipeline = FEPipeline(fe_enabled=True).fit(dm)
train_data = pipeline.transform(dm)
print(train_data)
print(train_data.data)
def model_fit(_id,obj,paramsj,X_trainj,y_trainj):
info_path = './models_information/'+_id+'_information'
info_file = open(info_path,'w')
print('Model training begins!')
try:
# read data
X_train = np.array(pd.DataFrame(json.loads(X_trainj)))
y_train = np.array(pd.DataFrame(json.loads(y_trainj)))[:,0]
params = json.loads(paramsj)
#print(y_train)
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
save_dir = '../data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# train mode
if(obj == 'clf'):
mdl = Classifier(time_limit=params['time_limit'],
output_dir=save_dir,
ensemble_method=params['ensemble_method'],
evaluation=params['evaluation'],
metric=params['metric'],
n_jobs=4)
elif(obj == 'reg'):
mdl = rgs = Regressor(metric=params['metric'],
ensemble_method=params['ensemble_method'],
evaluation=params['evaluation'],
time_limit=params['time_limit'],
output_dir=save_dir,
random_state=1,
n_jobs=n_jobs)
mdl.fit(train_data)
except:
print('Model training failed!')
info_file.write('Model training failed!')
info_file.close()
return -1
result = dict()
result['best_algo_id'] = str(mdl.best_algo_id)
result['best_hpo_config'] = str(mdl.best_hpo_config)
result['nbest_algo_id'] = str(mdl.nbest_algo_id)
result['best_perf'] = str(mdl.best_perf)
result['best_fe_config'] = str(mdl.best_fe_config)
result['get_ens_model_info'] = str(mdl.get_ens_model_info)
#get_ens_model_info is not realized in this version yet
info_file.write(json.dumps(result))
info_file.close()
print('Model training finished!')
return 0
def fetch_data(task_id):
dm = DataManager()
train_data_path = data_dir + 'data/p%d.csv' % task_id
test_data_path = data_dir + 'data/test_data.csv'
if not os.path.exists(train_data_path) or not os.path.exists(test_data_path):
create_csv(task_id)
train_data_node = dm.load_train_csv(train_data_path, label_col=-1, header='infer', sep=',')
print('loading train data finished.')
test_data_node = dm.load_test_csv(test_data_path, has_label=False, header='infer', sep=',')
print('loading test data finished.')
return train_data_node, test_data_node
def load_tabular_data(self, data_path):
self.data_manager = DataManager()
train_data_node = self.data_manager.load_train_csv(
data_path,
label_col=self.label_column,
header=self.header,
sep=self.sep,
na_values=list(self.nan_values))
task_type = REGRESSION if self.is_regression else CLASSIFICATION
self._process_pipeline = FEPipeline(fe_enabled=False,
metric='acc',
task_type=task_type)
return self._process_pipeline.fit_transform(train_data_node)
class TabularDataset(BaseDataset):
def __init__(self,
data_path: str,
is_regression=False,
label_column=-1,
header='infer',
sep=',',
nan_values=("n/a", "na", "--", "-", "?"),
train_val_split: bool = False,
val_split_size: float = 0.2):
super().__init__()
self.is_regression = is_regression
self.train_val_split = train_val_split
self.val_split_size = val_split_size
self.data_path = data_path
self.label_column = label_column
self.header = header
self.sep = sep
self.nan_values = nan_values
self.data_manager = None
self._process_pipeline = None
def load_tabular_data(self, data_path):
self.data_manager = DataManager()
train_data_node = self.data_manager.load_train_csv(
data_path,
label_col=self.label_column,
header=self.header,
sep=self.sep,
na_values=list(self.nan_values))
task_type = REGRESSION if self.is_regression else CLASSIFICATION
self._process_pipeline = FEPipeline(fe_enabled=False,
metric='acc',
task_type=task_type)
return self._process_pipeline.fit_transform(train_data_node)
def load_data(self):
self.train_dataset = self.load_tabular_data(self.data_path)
def load_test_data(self):
test_data_node = self.data_manager.load_test_csv(self.test_data_path,
has_label=False,
keep_default_na=True,
header=self.header,
sep=self.sep)
self.test_dataset = self._process_pipeline.transform(test_data_node)
def post(self):
start_time = time.time()
args = self.parser.parse_args()
_id = request.form['model_name']
X_file = request.files['data_file_X']
y_file = request.files['data_file_y']
_id = request.form['model_name']
obj = request.form['objective']
# read data
X_train = np.array(pd.read_csv(X_file))
y_train = np.array(pd.read_csv(y_file))[:, 0]
print(y_train)
if not (obj):
obj = 'clf'
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# train mode
if (obj == 'clf'):
mdl = Classifier(time_limit=100,
output_dir=save_dir,
ensemble_method='bagging',
evaluation='holdout',
metric='acc',
n_jobs=4)
elif (obj == 'reg'):
mdl = rgs = Regressor(metric='mse',
ensemble_method=ensemble_method,
evaluation=eval_type,
time_limit=time_limit,
output_dir=save_dir,
random_state=1,
n_jobs=n_jobs)
model_fit(_id, mdl, train_data)
self.model_factory.add_pipeline(mdl, train_data, _id)
print(self.model_factory)
result = {
'trainTime': time.time() - start_time,
'trainShape': X_train.shape
}
self.model_factory[params['pipeline_id']]['stats'] = result
return json.dumps(result)
def post(self):
args = self.parser.parse_args()
X_file = request.files['data_file_X']
X_test = np.array(pd.read_csv(X_file))
y_test = np.zeros(X_test.shape[0])
dm = DataManager(X_test, y_test)
test_data = dm.get_data_node(X_test, y_test)
_id = request.form['model_name']
proba = request.form['need_proba']
mdl = self.model_factory.pipelines[_id]['model']
if (proba):
try:
y_pred = mdl.predict_proba(test_data)
except:
y_pred = mdl.predict(test_data)
else:
y_pred = mdl.predict(test_data)
print(y_pred)
#params = read_params(args['params'].stream)
return pd.DataFrame(y_pred).to_json()
def load_data(dataset, data_dir='./'):
dm = DataManager(na_values=[])
data_path = data_dir + 'data/openml100/%s.csv' % dataset
# Load train data.
if dataset in ['higgs', 'cjs', 'Australian', 'monks-problems-1', 'monks-problems-2', 'monks-problems-3', 'profb',
'JapaneseVowels']:
label_column = 0
else:
label_column = -1
header = 'infer'
sep = ','
train_data_node = dm.load_train_csv(data_path, label_col=label_column, header=header, sep=sep,
na_values=["n/a", "na", "--", "?"], keep_default_na=False)
train_data_node = dm.preprocess_fit(train_data_node)
X, y = train_data_node.data
feature_types = train_data_node.feature_types
return X, y, feature_types
def evaluate_data_manager():
# from data_manager import DataManager
# dm = DataManager()
# train_df = dm.load_train_csv("data/proprocess_data.csv")
# print(train_df)
# print(dm.feature_types)
# print(dm.missing_flags)
from solnml.utils.data_manager import DataManager
import numpy as np
X = np.array([[1, 2, 3, 4], [1, 'asfd', 2, 1.4]])
y = [1, 2]
dm = DataManager(X, y)
print(dm.feature_types)
print(dm.missing_flags)
class Classifier():
def __init__(self,
dataset_name='default_dataset_name',
time_limit=10800,
amount_of_resource=None,
metric='acc',
include_algorithms=None,
enable_meta_algorithm_selection=True,
ensemble_method='ensemble_selection',
ensemble_size=50,
per_run_time_limit=150,
random_state=1,
n_jobs=1,
evaluation='holdout',
impute=False,
impute_method=['MatrixFactorization', 'KNN', 'IterativeSVD'],
pre_fs=True,
fb_k=300,
fb_r0=50,
fb_max_iter=50,
fb_population_size=10,
fb_n0=500,
fb_metric='accuracy',
output_dir="/tmp/"):
self.dataset_name = dataset_name
self.metric = metric
self.task_type = None
self.time_limit = time_limit
self.amount_of_resource = amount_of_resource
self.include_algorithms = include_algorithms
self.enable_meta_algorithm_selection = enable_meta_algorithm_selection
self.ensemble_method = ensemble_method
self.ensemble_size = ensemble_size
self.per_run_time_limit = per_run_time_limit
self.random_state = random_state
self.n_jobs = n_jobs
self.evaluation = evaluation
self.output_dir = output_dir
self.dm = None
#imputation arguments
self.impute = impute
self.impute_method = impute_method
self.impute_operator = {}
#fb arguments
self.pre_fs = pre_fs
self.fb_k = fb_k
self.fb_r0 = fb_r0
self.fb_max_iter = fb_max_iter
self.fb_population_size = fb_population_size
self.fb_n0 = fb_n0
self.fb_metric = fb_metric #['accuracy','f1_score']
self.origin_X = None
self.origin_y = None
self.fb_operator = None
self.headers = None
self.selected_headers = None
self.train_data = {}
self.test_data = {}
self.mdl = {}
self.y_pred = {}
# Create output directory.
if not os.path.exists(output_dir):
os.makedirs(output_dir)
if not self.impute:
self.impute_method = ['mean'] #simply impute anyway
def run_impute(self, X, state='train'):
if state == 'train':
self.train_data['ave'] = np.zeros([X.shape[0], X.shape[1]])
for imp_method in self.impute_method:
print('Impute ' + imp_method + ' starts!')
if imp_method == 'mean':
imp_ope = SimpleFill()
if imp_method == 'KNN':
imp_ope = KNN()
if imp_method == 'IterativeSVD':
imp_ope = IterativeSVD()
if imp_method == 'MatrixFactorization':
imp_ope = MatrixFactorization()
X_filled = imp_ope.fit_transform(X)
self.train_data[imp_method] = X_filled
self.impute_operator[imp_method] = imp_ope
self.train_data['ave'] += X_filled
print('Impute ' + imp_method + ' ends!')
self.train_data['ave'] /= len(self.impute_method)
return 0
def feature_selection(self, X, y, state='train'):
if state == 'train':
print('Feature_selection starts!')
fb = FeatureBand(r0=self.fb_r0,
n0=self.fb_n0,
clf=load_clf('logistic'),
max_iter=self.fb_max_iter,
k=self.fb_k,
population_size=self.fb_population_size,
local_search=True)
times, iter_best, global_best = fb.fit(X,
y,
metrics=self.fb_metric)
for key in self.train_data:
self.train_data[key] = fb.transform(self.train_data[key])
self.fb_operator = fb
if self.headers is not None:
self.selected_headers = self.headers[fb.featrue_selected]
return 0
def preprocess(self, X, y):
self.run_impute(X)
if X.shape[1] > self.fb_k and self.pre_fs == True:
self.feature_selection(self.train_data['ave'], y)
return 0
def fit(self, X, y):
if isinstance(X, pd.DataFrame):
self.headers = X.columns
X = np.array(X)
y = np.array(y)
self.origin_X = X
self.origin_y = y
if (np.sum(np.isnan(X)) == 0):
self.impute = False
self.impute_method = ['mean']
self.preprocess(X, y)
for key in self.impute_operator:
X_train = self.train_data[key]
self.dm = DataManager(X_train, y)
train_data = self.dm.get_data_node(X_train, y)
self.mdl[key] = soln_Classifier(
time_limit=self.time_limit / len(self.impute_method),
output_dir=self.output_dir,
ensemble_method=self.ensemble_method,
evaluation=self.evaluation,
metric=self.metric,
n_jobs=self.n_jobs)
self.mdl[key].fit(train_data)
return 0
def predict_proba(self, X_test):
y_pred = None
X_test = np.array(X_test)
for key in self.impute_operator:
if np.sum(np.isnan(X_test)) > 0:
X_test_filled = self.impute_operator[key].fit_transform(X_test)
else:
X_test_filled = X_test
if X_test.shape[1] > self.fb_k and self.pre_fs == True:
X_test_filled = self.fb_operator.transform(X_test_filled)
test_data = self.dm.get_data_node(X_test_filled, [])
self.y_pred[key] = self.mdl[key].predict_proba(test_data)
if y_pred is None:
y_pred = self.y_pred[key]
else:
y_pred += self.y_pred[key]
y_pred /= len(self.impute_operator)
return y_pred
def predict(self, X_test):
return np.argmax(self.predict_proba(X_test), axis=1)
@property
def get_feature_selected(self):
if self.fb_operator is not None:
return self.fb_operator.featrue_selected
def feature_analysis(self, topk=30):
from lightgbm import LGBMClassifier
relation_list = {}
result = None
importance_array = []
for key in self.impute_operator:
mdl = self.mdl[key]
data = self.dm.get_data_node(self.train_data[key], self.origin_y)
data_tf = mdl.data_transform(data).data[0]
sub_topk = min(int(topk / len(self.impute_operator)),
data_tf.shape[1])
lgb = LGBMClassifier()
lgb.fit(data_tf, self.origin_y)
_importance = lgb.feature_importances_
index_needed = np.argsort(-_importance)[:sub_topk]
temp_array = _importance[index_needed]
temp_array = temp_array / np.max(temp_array)
importance_array.append(temp_array)
relation_list[key] = mdl.feature_corelation(data, index_needed)
relation_list[key].index = key + relation_list[key].index
if self.selected_headers is not None:
relation_list[key].columns = list(self.selected_headers)
elif self.fb_operator is not None:
relation_list[key].columns = [
'origin_fearure' + str(it)
for it in self.fb_operator.featrue_selected
]
if result is None:
result = relation_list[key]
else:
result = result.append(relation_list[key])
importance_frame = pd.DataFrame(np.hstack(importance_array))
importance_frame.columns = ['feature_importance']
importance_frame.index = result.index
return pd.concat([importance_frame, result], axis=1)
return result
def save(self, save_dir):
saveloadmodel.save(self, save_dir, task_type='CLF')
args = parser.parse_args()
time_limit = args.time_limit
eval_type = args.eval_type
n_jobs = args.n_jobs
ensemble_method = args.ens_method
if ensemble_method == 'none':
ensemble_method = None
print('==> Start to evaluate with Budget %d' % time_limit)
boston = load_boston()
X, y = boston.data, boston.target
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=1)
dm = DataManager(X_train, y_train)
train_data = dm.get_data_node(X_train, y_train)
test_data = dm.get_data_node(X_test, y_test)
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):
os.makedirs(save_dir)
rgs = Regressor(metric='mse',
dataset_name='boston',
ensemble_method=ensemble_method,
evaluation=eval_type,
time_limit=time_limit,
output_dir=save_dir,
random_state=1,
n_jobs=n_jobs)
default='ensemble_selection',
choices=['none', 'bagging', 'blending', 'stacking', 'ensemble_selection'])
parser.add_argument('--n_jobs', type=int, default=1)
args = parser.parse_args()
time_limit = args.time_limit
eval_type = args.eval_type
n_jobs = args.n_jobs
ensemble_method = args.ens_method
if ensemble_method == 'none':
ensemble_method = None
print('==> Start to evaluate with Budget %d' % time_limit)
dm = DataManager()
train_node = dm.load_train_csv("train_dataset.csv",
label_col=-1,
header='infer',
na_values=['nan', '?'])
test_node = dm.load_test_csv("test_dataset.csv",
header='infer',
has_label=True)
from solnml.components.utils.constants import REGRESSION
pipeline = FEPipeline(fe_enabled=False, task_type=REGRESSION)
train_data = pipeline.fit_transform(train_node)
test_data = pipeline.transform(test_node)
save_dir = './data/eval_exps/soln-ml'
if not os.path.exists(save_dir):