def test_evalAEF5():
evaluation = definedEvaluation()
e = evaluation.eval_function("AEF5_0")
value0 = e(bin_pred, dtrain_cont)
evaluation = definedEvaluation()
e = evaluation.eval_function("AEF5")
value1 = e(bin_pred, dtrain_bin)
assert value0 == value1
assert np.round([value0[1]], 2) == 0.99
def test_evalefr015():
evaluation = definedEvaluation()
e = evaluation.eval_function("EFR015_0")
value0 = e(bin_pred, dtrain_cont)
evaluation = definedEvaluation()
e = evaluation.eval_function("EFR015")
value1 = e(bin_pred, dtrain_bin)
assert value0 == value1
assert np.round([value0[1]], 2) == 1.34
def test_evalNEFauc25():
evaluation = definedEvaluation()
e = evaluation.eval_function("NEFAUC25_0")
value0 = e(bin_pred, dtrain_cont)
evaluation = definedEvaluation()
e = evaluation.eval_function("NEFAUC25")
value1 = e(bin_pred, dtrain_bin)
assert value0 == value1
assert np.round([value0[1]], 2) == 0.55
def test_evalprauc():
evaluation = definedEvaluation()
e = evaluation.eval_function("PRAUC_0")
value0 = e(bin_pred, dtrain_cont)
evaluation = definedEvaluation()
e = evaluation.eval_function("PRAUC")
value1 = e(bin_pred, dtrain_bin)
assert value0 == value1
assert np.round([value0[1]], 2) == 0.51
def test_evalReliabilityScore():
evaluation = definedEvaluation()
e = evaluation.eval_function("ReliabilityScore_0")
value0 = e(bin_pred, dtrain_cont)
evaluation = definedEvaluation()
e = evaluation.eval_function("ReliabilityScore")
value1 = e(bin_pred, dtrain_bin)
assert value0 == value1
assert np.round([value0[1]], 2) == 0.25
def test_evalLogloss():
evaluation = definedEvaluation()
e = evaluation.eval_function("Logloss_0")
value0 = e(bin_pred, dtrain_cont)
evaluation = definedEvaluation()
e = evaluation.eval_function("Logloss")
value1 = e(bin_pred, dtrain_bin)
assert value0 == value1
assert np.round([value0[1]], 2) == 1504.69
def __init__(self,xgbData,eval_name,model_type,model_name):
"""
Parameters:
-----------
xgbData: object
Default data object that contains training data, testing data, cv-fold
info, and label.
eval_name: str
Name of evaluation metric used to monitor training process. Must in
pre-defined evaluation list.
model_type: str
Name of model type you want to use. Must in pre-defined model type list.
model_name: str
Unique name for this model.
"""
self.name = model_name
self.__preDefined_model = defined_model.definedModel()
self.__DEFINED_MODEL_TYPE = self.__preDefined_model.model_type()
self.__preDefined_eval = defined_eval.definedEvaluation()
self.__DEFINED_EVAL = self.__preDefined_eval.eval_list()
self.__xgbData = xgbData
self.__preDefined_eval.validate_eval_name(eval_name)
self.__eval_name = eval_name
self.__preDefined_model.validate_model_type(model_type)
self.__model_type_writeout = model_type
self.__collect_model = None
self.__track_best_ntree = pd.DataFrame(columns = ['model_name','best_ntree'])
self.__best_score = list()
self.__param = self.__preDefined_model.model_param(model_type)
self.__eval_function = self.__preDefined_eval.eval_function(self.__eval_name)
self.__MAXIMIZE = self.__preDefined_eval.is_maximize(self.__eval_name)
self.__STOPPING_ROUND = self.__preDefined_eval.stopping_round(self.__eval_name)
self.__holdout = None
def test_defined_eval():
eval = defined_eval.definedEvaluation()
assert eval.is_maximize('ROCAUC') == True
assert eval.stopping_round('ROCAUC') == 100
mark = 0
try:
eval.is_maximize('not_exist_eval_name')
assert mark == 1
except ValueError:
mark = 1
def __init__(self,xgbData,list_firstLayerModel,eval_name,model_type,model_name):
"""Use holdout(out of fold) predictions from several firstLayerModels as
training features to train a secondLayerModel.(So called stacking model)
Parameters:
-----------
xgbData: object
contains the label you want to use in second layer model.
list_firstLayerModel: list
list contains firstLayerModel.
eval_name: str
Name of evaluation metric used to monitor training process. Must in
pre-defined evaluation list.
model_type: str
Name of model type you want to use. Must in pre-defined model type list.
model_name: str
Unique name for this model.
"""
self.name = model_name
self.__preDefined_model = defined_model.definedModel()
self.__preDefined_eval = defined_eval.definedEvaluation()
self.__DEFINED_EVAL = self.__preDefined_eval.eval_list()
self.__xgbData = xgbData
assert all([isinstance(item,first_layer_model.firstLayerModel) for item in list_firstLayerModel])
self.__list_firstLayerModel = list_firstLayerModel
self.__preDefined_eval.validate_eval_name(eval_name)
self.__eval_name = eval_name
self.__preDefined_model.validate_model_type(model_type)
self.__model_type_writeout = model_type
self.__collect_model = None
self.__track_best_ntree = pd.DataFrame(columns = ['model_name','best_ntree'])
self.__best_score = list()
self.__firstLayerModel_prediction = None
self.__param = self.__preDefined_model.model_param(model_type)
self.__eval_function = self.__preDefined_eval.eval_function(self.__eval_name)
self.__MAXIMIZE = self.__preDefined_eval.is_maximize(self.__eval_name)
self.__STOPPING_ROUND = self.__preDefined_eval.stopping_round(self.__eval_name)
self.__holdout = None
# Read command line inpu and match input.
with open(sys.argv[1], 'r') as f:
info = f.read()
info = pd.read_json(info)
target_name = np.str(info.loc['target_name'][0])
dir_train = np.str(info.loc['full_directory_to_training_data'][0])
dir_test = np.str(info.loc['full_directory_to_dataToPredict_if_exit'][0])
smile_colname = np.str(info.loc['smile_column_name'][0])
label_name_list = info.loc['label_name_list'][0]
label_name_list = [np.str(item) for item in label_name_list]
eval_name = np.str(info.loc['evaluation_name'][0])
dir_to_store = np.str(info.loc['full_directory_to_store_prediction'][0])
maccKeys_column_name = np.str(info.loc['maccKeys_column_name'][0])
ecfp1024_column_name = np.str(info.loc['ecfp1024_column_name'][0])
preDefined_eval = defined_eval.definedEvaluation()
preDefined_eval.validate_eval_name(eval_name)
df = pd.read_csv(dir_train)
# identify NA row.
missing_row = pd.isnull(df.loc[:, label_name_list[0]])
df = df.loc[~missing_row]
df = df.reset_index(drop=True)
print 'Preparing training data fingerprints'
# morgan(ecfp) fp
morgan_fp = df.copy()
morgan_fp = morgan_fp.rename(columns={'ecfp1024': 'fingerprint'})
# MACCSkeys fp
maccs_fp = df.copy()
maccs_fp = maccs_fp.rename(columns={'maccKeys': 'fingerprint'})
comb1 = (morgan_fp, label_name_list)
comb2 = (maccs_fp, label_name_list)
def train(self):
"""
Train the model. Train and check potential first and second layer models.
"""
evaluation_metric_name = self.__eval_name
print 'Building first layer models'
#---------------------------------first layer models ----------
for data_dict in self.__setting_list:
for model_type in data_dict['model_type']:
unique_name = 'layer1_' + data_dict['data_name'] + '_' + model_type + '_' + evaluation_metric_name
model = first_layer_model.firstLayerModel(data_dict['data'],
evaluation_metric_name,model_type,unique_name)
# Retrieve default parameter and change default seed.
default_param,default_MAXIMIZE,default_STOPPING_ROUND = model.get_param()
default_param['seed'] = self.seed
if self.__verbose == True:
default_param['silent'] = 1
elif self.__verbose == False:
default_param['verbose_eval'] = False
model.update_param(default_param,default_MAXIMIZE,default_STOPPING_ROUND)
model.xgb_cv()
model.generate_holdout_pred()
self.__layer1_model_list.append(model)
#------------------------------------second layer models
layer2_label_data = self.__setting_list[0]['data'] # layer1 data object containing the label for layer2 model
layer2_modeltype = ['GbtreeLogistic','GblinearLogistic']
layer2_evaluation_metric_name = [self.__eval_name]
print 'Building second layer models'
for evaluation_metric_name in layer2_evaluation_metric_name:
for model_type in layer2_modeltype:
unique_name = 'layer2' + '_' + model_type + '_' + evaluation_metric_name
l2model = second_layer_model.secondLayerModel(layer2_label_data,self.__layer1_model_list,
evaluation_metric_name,model_type,unique_name)
l2model.second_layer_data()
# Retrieve default parameter and change default seed.
default_param,default_MAXIMIZE,default_STOPPING_ROUND = l2model.get_param()
default_param['seed'] = self.seed
if self.__verbose == True:
default_param['silent'] = 0
elif self.__verbose == False:
default_param['verbose_eval'] = False
l2model.update_param(default_param,default_MAXIMIZE,default_STOPPING_ROUND)
l2model.xgb_cv()
self.__layer2_model_list.append(l2model)
#------------------------------------ evaluate model performance on test data
# prepare test data, retrive from layer1 data
list_TestData = []
for data_dict in self.__setting_list:
for model_type in data_dict['model_type']:
list_TestData.append(data_dict['data'].get_dtest())
test_label = layer2_label_data.get_testLabel()
test_result_list = []
i = 0
for evaluation_metric_name in layer2_evaluation_metric_name:
for model_type in layer2_modeltype:
test_result = eval_testset.eval_testset(self.__layer2_model_list[i],
list_TestData,test_label,
evaluation_metric_name)
test_result_list.append(test_result)
i += 1
# merge cv and test result together. Calcuate the weighted average of
# cv and test result for each model(layer1, layer2 model). Then use the best
# model to predict.
all_model = self.__layer1_model_list + self.__layer2_model_list
result = []
for model in all_model:
result = result + [item for item in np.array(model.cv_score_df())[0]]
# Retrieve corresponding name of cv result
result_index = []
for model in all_model:
result_index.append(model.name)
# create a dataframe
cv_result = pd.DataFrame({'cv_result' : result},index = result_index)
test_result = pd.concat(test_result_list,axis = 0,ignore_index=False)
test_result = test_result.rename(columns = {self.__eval_name:'test_result'})
#selet distinct row.
test_result['temp_name'] = test_result.index
test_result = test_result.drop_duplicates(['temp_name'])
test_result = test_result.drop('temp_name',1)
cv_test = pd.merge(cv_result,test_result,how='left',left_index=True,right_index=True)
self.__num_folds = np.float64(self.__num_folds)
cv_test['weighted_score'] = cv_test.cv_result * (self.__num_folds-1)/self.__num_folds + cv_test.test_result * (1/self.__num_folds)
weighted_score = cv_test.cv_result * (self.__num_folds-1)/self.__num_folds + cv_test.test_result * (1/self.__num_folds)
# Determine does current evaluation metric need to maximize or minimize
eval_info = defined_eval.definedEvaluation()
is_max = eval_info.is_maximize(self.__eval_name)
if is_max:
position = np.where(cv_test.weighted_score == cv_test.weighted_score.max())
best_model_name = cv_test.weighted_score.iloc[position].index[0]
else:
position = np.where(cv_test.weighted_score == cv_test.weighted_score.min())
best_model_name = cv_test.weighted_score.iloc[position].index[0]
# find best model
all_model_name = [model.name for model in all_model]
model_position = all_model_name.index(best_model_name)
self.__best_model = all_model[model_position]
self.__best_model_result = pd.DataFrame(cv_test.loc[self.__best_model.name])
self.__all_model_result = cv_test
def __prepare_result(self):
# merge cv and test result together. Calcuate the weighted average of
# cv and test result for each model(layer1, layer2 model). Then use the best
# model to predict.
all_model = self.__layer1_model_list + self.__layer2_model_list
result = []
for model in all_model:
result = result + [item for item in np.array(model.cv_score_df())[0]]
# Retrieve corresponding name of cv result
result_index = []
for model in all_model:
result_index.append(model.name)
# create a dataframe
cv_result = pd.DataFrame({'cv_result' : result},index = result_index)
#------------------------------------ evaluate model performance on test data
# prepare test data, retrive from layer1 data
if self.__createTestset:
list_TestData = []
layer2_modeltype = self.__layer2_modeltype
for data_dict in self.__setting_list:
for model_type in data_dict['model_type']:
list_TestData.append(data_dict['data'].get_dtest())
test_label = layer2_label_data.get_testLabel()
test_result_list = []
i = 0
for evaluation_metric_name in layer2_evaluation_metric_name:
for model_type in layer2_modeltype:
test_result = eval_testset.eval_testset(self.__layer2_model_list[i],
list_TestData,test_label,
evaluation_metric_name)
test_result_list.append(test_result)
i += 1
test_result = pd.concat(test_result_list,axis = 0,ignore_index=False)
test_result = test_result.rename(columns = {self.__eval_name:'test_result'})
#selet distinct row.
test_result['temp_name'] = test_result.index
test_result = test_result.drop_duplicates(['temp_name'])
test_result = test_result.drop('temp_name',1)
cv_test = pd.merge(cv_result,test_result,how='left',left_index=True,right_index=True)
self.__num_folds = np.float64(self.__num_folds)
cv_test['weighted_score'] = cv_test.cv_result * (self.__num_folds-1)/self.__num_folds + cv_test.test_result * (1/self.__num_folds)
weighted_score = cv_test.cv_result * (self.__num_folds-1)/self.__num_folds + cv_test.test_result * (1/self.__num_folds)
else:
cv_test = cv_result
cv_test['weighted_score'] = cv_result.cv_result
# Based on user specific finalModel
if self.__finalModel == None:
final_cv_test = cv_test
else:
finalModel_names = [item for item in list(cv_test.index) if self.__finalModel in item]
final_cv_test = cv_test.loc[finalModel_names]
# Determine does current evaluation metric need to maximize or minimize
eval_info = defined_eval.definedEvaluation()
is_max = eval_info.is_maximize(self.__eval_name)
if is_max:
position = np.where(final_cv_test.weighted_score == final_cv_test.weighted_score.max())
best_model_name = final_cv_test.weighted_score.iloc[position].index[0]
else:
position = np.where(final_cv_test.weighted_score == final_cv_test.weighted_score.min())
best_model_name = final_cv_test.weighted_score.iloc[position].index[0]
# find best model
all_model_name = [model.name for model in all_model]
model_position = all_model_name.index(best_model_name)
self.__best_model = all_model[model_position]
self.__best_model_result = pd.DataFrame(cv_test.loc[self.__best_model.name])
self.__all_model_result = cv_test
# Find model contains the final label
if self.__final_labelType == 'binary':
model_has_finalLabel = [item for item in list(cv_test.index) if 'Logistic' in item]
model_position = all_model_name.index(model_has_finalLabel[0])
self.__model_has_finalLabel = all_model[model_position]
elif self.__final_labelType == 'continuous':
model_has_finalLabel = [item for item in list(cv_test.index) if 'Regression' in item]
model_position = all_model_name.index(model_has_finalLabel[0])
self.__model_has_finalLabel = all_model[model_position]
