# get only centrality selected
train_test_data_cent = [
pd.DataFrame(), [],
pd.DataFrame(), []
]
train_test_data_cent[0] = train_test_data[0].query(
f'matter {split_ineq_sign} and centrality > {cent_bins[0]} and centrality < {cent_bins[1]} and ct >= {ct_bins_df[0]} and ct < {ct_bins_df[1]}'
)
train_test_data_cent[2] = train_test_data[2].query(
f'matter {split_ineq_sign} and centrality > {cent_bins[0]} and centrality < {cent_bins[1]} and ct >= {ct_bins_df[0]} and ct < {ct_bins_df[1]}'
)
train_test_data_cent[1] = train_test_data_cent[0]['y_true']
train_test_data_cent[3] = train_test_data_cent[2]['y_true']
# get predictions for training and test sets
test_y_score = model_hdl.predict(train_test_data_cent[2])
train_y_score = model_hdl.predict(train_test_data_cent[0])
# second condition needed because of issue with Qt libraries
if MAKE_TRAIN_TEST_PLOT and not MAKE_PRESELECTION_EFFICIENCY:
if not os.path.isdir(f'{PLOT_DIR}/train_test_out'):
os.mkdir(f'{PLOT_DIR}/train_test_out')
plot_utils.plot_output_train_test(model_hdl,
train_test_data_cent,
logscale=True,
density=True,
labels=leg_labels)
plt.savefig(
f'{PLOT_DIR}/train_test_out/{bin_df}_out.pdf')
plot_utils.plot_feature_imp(train_test_data_cent[0],
# data preparation
DIGITS_DATA = datasets.load_digits(n_class=2)
DIGITS = pd.DataFrame(DIGITS_DATA.data[:, 0:10]) # pylint: disable=E1101
Y_DIGITS = DIGITS_DATA.target # pylint: disable=E1101
SIG_DF = DIGITS[Y_DIGITS == 1]
BKG_DF = DIGITS[Y_DIGITS == 0]
TRAIN_SET, TEST_SET, Y_TRAIN, Y_TEST = train_test_split(
DIGITS, Y_DIGITS, test_size=0.5, random_state=42)
DATA = [TRAIN_SET, Y_TRAIN, TEST_SET, Y_TEST]
# --------------------------------------------
# training and testing
INPUT_MODEL = xgb.XGBClassifier()
MODEL = ModelHandler(INPUT_MODEL)
MODEL.train_test_model(DATA)
Y_PRED = MODEL.predict(DATA[2])
Y_PRED_TRAIN = MODEL.predict(DATA[0])
EFFICIENCY, THRESHOLD = analysis_utils.bdt_efficiency_array(DATA[3], Y_PRED, n_points=10)
# --------------------------------------------
def test_plot_distr():
"""
Test the feature distribution plot
"""
assert isinstance(plot_utils.plot_distr(
[SIG_DF, BKG_DF], SIG_DF.columns), np.ndarray)
def test_plot_corr():
"""
def train_test(inputCfg, PtBin, OutPutDirPt, TrainTestData, iBin): #pylint: disable=too-many-statements, too-many-branches
'''
function for model training and testing
'''
n_classes = len(np.unique(TrainTestData[3]))
modelClf = xgb.XGBClassifier(use_label_encoder=False)
TrainCols = inputCfg['ml']['training_columns']
HyperPars = inputCfg['ml']['hyper_par'][iBin]
if not isinstance(TrainCols, list):
print('\033[91mERROR: training columns must be defined!\033[0m')
sys.exit()
if not isinstance(HyperPars, dict):
print(
'\033[91mERROR: hyper-parameters must be defined or be an empty dict!\033[0m'
)
sys.exit()
ModelHandl = ModelHandler(modelClf, TrainCols, HyperPars)
# hyperparams optimization
if inputCfg['ml']['hyper_par_opt']['do_hyp_opt']:
print('Perform bayesian optimization')
BayesOptConfig = inputCfg['ml']['hyper_par_opt']['bayes_opt_config']
if not isinstance(BayesOptConfig, dict):
print('\033[91mERROR: bayes_opt_config must be defined!\033[0m')
sys.exit()
if n_classes > 2:
average_method = inputCfg['ml']['roc_auc_average']
roc_method = inputCfg['ml']['roc_auc_approach']
if not (average_method in ['macro', 'weighted']
and roc_method in ['ovo', 'ovr']):
print(
'\033[91mERROR: selected ROC configuration is not valid!\033[0m'
)
sys.exit()
if average_method == 'weighted':
metric = f'roc_auc_{roc_method}_{average_method}'
else:
metric = f'roc_auc_{roc_method}'
else:
metric = 'roc_auc'
print('Performing hyper-parameters optimisation: ...', end='\r')
OutFileHypPars = open(
f'{OutPutDirPt}/HyperParOpt_pT_{PtBin[0]}_{PtBin[1]}.txt', 'wt')
sys.stdout = OutFileHypPars
ModelHandl.optimize_params_bayes(
TrainTestData,
BayesOptConfig,
metric,
nfold=inputCfg['ml']['hyper_par_opt']['nfolds'],
init_points=inputCfg['ml']['hyper_par_opt']['initpoints'],
n_iter=inputCfg['ml']['hyper_par_opt']['niter'],
njobs=inputCfg['ml']['hyper_par_opt']['njobs'])
OutFileHypPars.close()
sys.stdout = sys.__stdout__
print('Performing hyper-parameters optimisation: Done!')
print(
f'Output saved in {OutPutDirPt}/HyperParOpt_pT_{PtBin[0]}_{PtBin[1]}.txt'
)
print(f'Best hyper-parameters:\n{ModelHandl.get_model_params()}')
else:
ModelHandl.set_model_params(HyperPars)
# train and test the model with the updated hyper-parameters
yPredTest = ModelHandl.train_test_model(
TrainTestData,
True,
output_margin=inputCfg['ml']['raw_output'],
average=inputCfg['ml']['roc_auc_average'],
multi_class_opt=inputCfg['ml']['roc_auc_approach'])
yPredTrain = ModelHandl.predict(TrainTestData[0],
inputCfg['ml']['raw_output'])
# save model handler in pickle
ModelHandl.dump_model_handler(
f'{OutPutDirPt}/ModelHandler_pT_{PtBin[0]}_{PtBin[1]}.pickle')
ModelHandl.dump_original_model(
f'{OutPutDirPt}/XGBoostModel_pT_{PtBin[0]}_{PtBin[1]}.model', True)
#plots
LegLabels = [
inputCfg['output']['leg_labels']['Bkg'],
inputCfg['output']['leg_labels']['Prompt']
]
if inputCfg['output']['leg_labels']['FD'] is not None:
LegLabels.append(inputCfg['output']['leg_labels']['FD'])
OutputLabels = [
inputCfg['output']['out_labels']['Bkg'],
inputCfg['output']['out_labels']['Prompt']
]
if inputCfg['output']['out_labels']['FD'] is not None:
OutputLabels.append(inputCfg['output']['out_labels']['FD'])
#_____________________________________________
plt.rcParams["figure.figsize"] = (10, 7)
MLOutputFig = plot_utils.plot_output_train_test(
ModelHandl,
TrainTestData,
80,
inputCfg['ml']['raw_output'],
LegLabels,
inputCfg['plots']['train_test_log'],
density=True)
if n_classes > 2:
for Fig, Lab in zip(MLOutputFig, OutputLabels):
Fig.savefig(
f'{OutPutDirPt}/MLOutputDistr{Lab}_pT_{PtBin[0]}_{PtBin[1]}.pdf'
)
else:
MLOutputFig.savefig(
f'{OutPutDirPt}/MLOutputDistr_pT_{PtBin[0]}_{PtBin[1]}.pdf')
#_____________________________________________
plt.rcParams["figure.figsize"] = (10, 9)
ROCCurveFig = plot_utils.plot_roc(TrainTestData[3], yPredTest, None,
LegLabels,
inputCfg['ml']['roc_auc_average'],
inputCfg['ml']['roc_auc_approach'])
ROCCurveFig.savefig(
f'{OutPutDirPt}/ROCCurveAll_pT_{PtBin[0]}_{PtBin[1]}.pdf')
pickle.dump(
ROCCurveFig,
open(f'{OutPutDirPt}/ROCCurveAll_pT_{PtBin[0]}_{PtBin[1]}.pkl', 'wb'))
#_____________________________________________
plt.rcParams["figure.figsize"] = (10, 9)
ROCCurveTTFig = plot_utils.plot_roc_train_test(
TrainTestData[3], yPredTest, TrainTestData[1], yPredTrain, None,
LegLabels, inputCfg['ml']['roc_auc_average'],
inputCfg['ml']['roc_auc_approach'])
ROCCurveTTFig.savefig(
f'{OutPutDirPt}/ROCCurveTrainTest_pT_{PtBin[0]}_{PtBin[1]}.pdf')
#_____________________________________________
PrecisionRecallFig = plot_utils.plot_precision_recall(
TrainTestData[3], yPredTest, LegLabels)
PrecisionRecallFig.savefig(
f'{OutPutDirPt}/PrecisionRecallAll_pT_{PtBin[0]}_{PtBin[1]}.pdf')
#_____________________________________________
plt.rcParams["figure.figsize"] = (12, 7)
FeaturesImportanceFig = plot_utils.plot_feature_imp(
TrainTestData[2][TrainCols], TrainTestData[3], ModelHandl, LegLabels)
n_plot = n_classes if n_classes > 2 else 1
for iFig, Fig in enumerate(FeaturesImportanceFig):
if iFig < n_plot:
label = OutputLabels[iFig] if n_classes > 2 else ''
Fig.savefig(
f'{OutPutDirPt}/FeatureImportance{label}_pT_{PtBin[0]}_{PtBin[1]}.pdf'
)
else:
Fig.savefig(
f'{OutPutDirPt}/FeatureImportanceAll_pT_{PtBin[0]}_{PtBin[1]}.pdf'
)
return ModelHandl
def benchmark_hyperparam_optimizers(filename_dict,
params,
params_range,
flag_dict,
presel_dict,
training_variables='',
testsize=0.75):
import time
from sklearn.metrics import roc_auc_score
N_run = 1
data_path = filename_dict['data_path']
analysis_path = filename_dict['analysis_path']
print('Loading MC signal')
mc_signal = TreeHandler()
mc_signal.get_handler_from_large_file(
file_name=data_path + filename_dict['MC_signal_filename'],
tree_name=filename_dict['MC_signal_table'])
print('MC signal loaded\n')
print('Loading background data for training')
background_ls = TreeHandler()
background_ls.get_handler_from_large_file(
file_name=data_path + filename_dict['train_bckg_filename'],
tree_name=filename_dict['train_bckg_table'])
background_ls.apply_preselections(presel_dict['train_bckg_presel'])
background_ls.shuffle_data_frame(size=min(background_ls.get_n_cand(),
mc_signal.get_n_cand() * 4))
print('Done\n')
train_test_data = train_test_generator([mc_signal, background_ls], [1, 0],
test_size=testsize)
if training_variables == '':
training_variables = train_test_data[0].columns.tolist()
model_clf = xgb.XGBClassifier()
model_hdl = ModelHandler(model_clf, training_variables)
times = []
roc = []
for i in range(N_run):
start = time.time()
model_hdl.optimize_params_bayes(train_test_data,
params_range,
'roc_auc',
njobs=-1)
model_hdl.train_test_model(train_test_data, )
y_pred_test = model_hdl.predict(
train_test_data[2], True) #used to evaluate model performance
roc.append(roc_auc_score(train_test_data[3], y_pred_test))
times.append(time.time() - start)
print('BAYES OPTIMIZATION WITH SKLEARN')
print('Mean time : ' + str(np.mean(time)))
print('Mean ROC : ' + str(np.mean(roc)))
print('--------------\n')
for i in range(N_run):
model_hdl.optimize_params_optuna(train_test_data,
params_range,
'roc_auc',
timeout=np.mean(times),
njobs=-1)
model_hdl.train_test_model(train_test_data, )
y_pred_test = model_hdl.predict(
train_test_data[2], True) #used to evaluate model performance
roc.append(roc_auc_score(train_test_data[3], y_pred_test))
print('OPTUNA')
print('Fixed time : ' + str(np.mean(time)))
print('Mean ROC : ' + str(np.mean(roc)))
print('--------------\n')
def train_test(inputCfg, PtMin, PtMax, OutPutDirPt, TrainTestData):
'''
function for model training and testing
'''
modelClf = xgb.XGBClassifier()
TrainCols = inputCfg['ml']['training_columns']
HyperPars = inputCfg['ml']['hyper_par']
if not isinstance(TrainCols, list):
print('ERROR: training columns must be defined!')
sys.exit()
if not isinstance(HyperPars, dict):
print('ERROR: hyper-parameters must be defined or be an empty dict!')
sys.exit()
ModelHandl = ModelHandler(modelClf, TrainCols, HyperPars)
# hyperparams optimization --> not working with multi-class classification at the moment
#HypRanges = {
# # # defines the maximum depth of a single tree (regularization)
# 'max_depth': (1, 30),
# 'learning_rate': (0.01, 0.3), # learning rate
# 'n_estimators': (50, 1000) # number of boosting trees
#}
#ModelHandl.optimize_params_bayes(TrainTestData, HypRanges, None)
# train and test the model with the updated hyperparameters
ModelHandl.train_test_model(TrainTestData)
yPredTest = ModelHandl.predict(TrainTestData[2],
inputCfg['ml']['raw_output'], True)
# save model handler in pickle
ModelHandl.dump_model_handler(
f'{OutPutDirPt}/ModelHandler_pT_{PtMin}_{PtMax}.pickle')
#plots
LegLabels = inputCfg['output']['leg_labels']
OutputLabels = inputCfg['output']['out_labels']
#_____________________________________________
plt.rcParams["figure.figsize"] = (10, 7)
MLOutputFig = plot_utils.plot_output_train_test(
ModelHandl,
TrainTestData,
80,
inputCfg['ml']['raw_output'],
LegLabels,
True,
inputCfg['plots']['train_test_log'],
density=True)
for Fig, Lab in zip(MLOutputFig, OutputLabels):
Fig.savefig(f'{OutPutDirPt}/MLOutputDistr{Lab}_pT_{PtMin}_{PtMax}.pdf')
#_____________________________________________
plt.rcParams["figure.figsize"] = (8, 7)
ROCCurveFig = plot_utils.plot_roc(TrainTestData[3], yPredTest, LegLabels)
ROCCurveFig.savefig(f'{OutPutDirPt}/ROCCurveAll_pT_{PtMin}_{PtMax}.pdf')
#_____________________________________________
PrecisionRecallFig = plot_utils.plot_precision_recall(
TrainTestData[3], yPredTest, LegLabels)
PrecisionRecallFig.savefig(
f'{OutPutDirPt}/PrecisionRecallAll_pT_{PtMin}_{PtMax}.pdf')
#_____________________________________________
plt.rcParams["figure.figsize"] = (12, 7)
FeaturesImportanceFig = plot_utils.plot_feature_imp(
TrainTestData[2][TrainCols], TrainTestData[3], ModelHandl)
for iFig, Fig in enumerate(FeaturesImportanceFig):
if iFig < 3:
Fig.savefig(
f'{OutPutDirPt}/FeatureImportance{OutputLabels[iFig]}_pT_{PtMin}_{PtMax}.pdf'
)
else:
Fig.savefig(
f'{OutPutDirPt}/FeatureImportanceAll_pT_{PtMin}_{PtMax}.pdf')
return ModelHandl
ct_bins_df_index = int(ct_bins[0]/5)
for ct_bins_df in zip(CT_BINS_APPLY[i_cent_bins][ct_bins_df_index][:-1], CT_BINS_APPLY[i_cent_bins][ct_bins_df_index][1:]):
bin_df = f'{split}_{cent_bins[0]}_{cent_bins[1]}_{ct_bins_df[0]}_{ct_bins_df[1]}'
print(bin_df)
# get only centrality selected
train_test_data_cent = [pd.DataFrame(), [], pd.DataFrame(), []]
train_test_data_cent[0] = train_test_data[0].query(f'matter {split_ineq_sign} and centrality > {cent_bins[0]} and centrality < {cent_bins[1]} and ct >= {ct_bins_df[0]} and ct < {ct_bins_df[1]}')
train_test_data_cent[2] = train_test_data[2].query(f'matter {split_ineq_sign} and centrality > {cent_bins[0]} and centrality < {cent_bins[1]} and ct >= {ct_bins_df[0]} and ct < {ct_bins_df[1]}')
train_test_data_cent[1] = train_test_data_cent[0]['y_true']
train_test_data_cent[3] = train_test_data_cent[2]['y_true']
# get predictions for training and test sets
print(train_test_data_cent[2])
if (ct_bins_df[0]==0):
continue
test_y_score = model_hdl.predict(train_test_data_cent[2], output_margin=False)
train_y_score = model_hdl.predict(train_test_data_cent[0], output_margin=False)
train_test_data_cent[0].loc[:,'model_output_prompt'] = train_y_score[:,2]
train_test_data_cent[2].loc[:,'model_output_prompt'] = test_y_score[:,2]
train_test_data_cent[0].loc[:,'model_output_non_prompt'] = train_y_score[:,1]
train_test_data_cent[2].loc[:,'model_output_non_prompt'] = test_y_score[:,1]
train_test_data_cent[0].loc[:,'model_output_background'] = train_y_score[:,0]
train_test_data_cent[2].loc[:,'model_output_background'] = test_y_score[:,0]
# write
train_test_data_cent[0].to_parquet(f'df/train_data_{cent_bins[0]}_{cent_bins[1]}_{ct_bins_df[0]}_{ct_bins_df[1]}.parquet.gzip',compression='gzip')
train_test_data_cent[2].to_parquet(f'df/test_data_{cent_bins[0]}_{cent_bins[1]}_{ct_bins_df[0]}_{ct_bins_df[1]}.parquet.gzip',compression='gzip')
# second condition needed because of issue with Qt libraries
if MAKE_TRAIN_TEST_PLOT and not MAKE_PRESELECTION_EFFICIENCY:
if not os.path.isdir(f'{PLOT_DIR}/train_test_out'):
def train_xgboost_model(signal,
background,
filename_dict,
params,
params_range,
flag_dict,
training_variables='',
testsize=0.5):
'''
Trains an XGBOOST model using hipe4ml and plot output distribution and feature importance
'''
print('Training XGBOOST model')
training_fig_path = filename_dict['analysis_path'] + "/images/training"
train_test_data = train_test_generator([signal, background], [1, 0],
test_size=testsize)
if training_variables == '':
training_variables = train_test_data[0].columns.tolist()
model_clf = xgb.XGBClassifier()
model_hdl = ModelHandler(model_clf, training_variables)
if not flag_dict['use_default_param']:
model_hdl.set_model_params(params)
if flag_dict['benchmark_opt']:
print('Benchamarking optimizers\n')
import time
from sklearn.metrics import roc_auc_score
times_sk = []
roc_sk = []
for i in range(1):
start = time.time()
model_hdl.optimize_params_bayes(train_test_data,
params_range,
'roc_auc',
njobs=-1)
model_hdl.train_test_model(train_test_data, )
y_pred_test = model_hdl.predict(
train_test_data[2], True) #used to evaluate model performance
roc_sk.append(roc_auc_score(train_test_data[3], y_pred_test))
times_sk.append(time.time() - start)
print('\nBAYES OPTIMIZATION WITH SKLEARN')
print('Mean time : ' + str(np.mean(times_sk)))
print('Mean ROC : ' + str(np.mean(roc_sk)))
print('--------------\n')
print('OPTUNA')
time = []
roc = []
for i in range(1):
for key in params:
if isinstance(params[key], str):
params_range[key] = params[key]
model_hdl.optimize_params_optuna(train_test_data,
params_range,
'roc_auc',
timeout=flag_dict['timeout'],
n_jobs=flag_dict['n_jobs'])
model_hdl.train_test_model(train_test_data, )
y_pred_test = model_hdl.predict(
train_test_data[2], True) #used to evaluate model performance
roc.append(roc_auc_score(train_test_data[3], y_pred_test))
print('\nBAYES OPTIMIZATION WITH SKLEARN')
print('Mean time : ' + str(np.mean(times_sk)))
print('Mean ROC : ' + str(np.mean(roc_sk)))
print('--------------\n')
print('OPTUNA')
print('Fixed time : ' + str(np.mean(time)))
print('Mean ROC : ' + str(np.mean(roc)))
print('--------------\n')
if flag_dict['optimize_bayes']:
import time
print('Doing Bayes optimization of hyperparameters\n')
start = time.time()
model_hdl.optimize_params_bayes(train_test_data,
params_range,
'roc_auc',
n_iter=700,
njobs=flag_dict['n_jobs'])
print('Elapsed time: ' + str(time.time() - start))
if flag_dict['optimize_optuna']:
print('Doing Optuna optimization of hyperparameters\n')
for key in params:
if isinstance(params[key], str):
params_range[key] = params[key]
study = model_hdl.optimize_params_optuna(train_test_data,
params_range,
scoring='roc_auc',
timeout=flag_dict['timeout'],
n_jobs=flag_dict['n_jobs'],
n_trials=None)
print('Parameters optimization done!\n')
if flag_dict['plot_optim']:
print('Saving optimization plots')
fig = optuna.visualization.plot_slice(study)
fig.write_image(training_fig_path + '/optuna_slice.png')
fig = optuna.visualization.plot_optimization_history(study)
fig.write_image(training_fig_path + '/optuna_history.png')
'''fig = optuna.visualization.plot_param_importances(study)
fig.write_image(training_fig_path + '/optuna_param_importance.png')
fig = optuna.visualization.plot_contour(study)
fig.write_image(training_fig_path + '/optuna_contour.png')'''
print('Done\n')
import joblib
joblib.dump(study, filename_dict['analysis_path'] + "model/study.pkl")
model_hdl.train_test_model(train_test_data, )
print(model_hdl.get_model_params())
print('Predicting values on training and test datas')
y_pred_train = model_hdl.predict(train_test_data[0], True)
y_pred_test = model_hdl.predict(train_test_data[2],
True) #used to evaluate model performance
print('Prediction done\n')
plt.rcParams["figure.figsize"] = (10, 7)
leg_labels = ['background', 'signal']
print('Saving Output comparison plot')
plt.figure()
ml_out_fig = plot_utils.plot_output_train_test(model_hdl,
train_test_data,
100,
True,
leg_labels,
True,
density=False)
plt.savefig(training_fig_path + '/output_train_test.png',
dpi=300,
facecolor='white')
plt.close()
print('Done\n')
print('Saving ROC AUC plot')
plt.figure()
roc_train_test_fig = plot_utils.plot_roc_train_test(
train_test_data[3], y_pred_test, train_test_data[1], y_pred_train,
None, leg_labels) #ROC AUC plot
plt.savefig(training_fig_path + '/ROC_AUC_train_test.png',
dpi=300,
facecolor='white')
import pickle
with open(training_fig_path + '/ROC_AUC_train_test.pickle', 'wb') as f:
pickle.dump(roc_train_test_fig, f)
plt.close()
print('Done\n')
print('Saving feature importance plots')
plt.figure()
feat_imp_1, feat_imp_2 = plot_utils.plot_feature_imp(train_test_data[2],
train_test_data[3],
model_hdl,
approximate=True)
feat_imp_1.savefig(training_fig_path +
'/feature_importance_HIPE4ML_violin.png',
dpi=300,
facecolor='white')
feat_imp_2.savefig(training_fig_path +
'/feature_importance_HIPE4ML_bar.png',
dpi=300,
facecolor='white')
plt.close()
print('Done\n')
efficiency_score_conversion(train_test_data, y_pred_test, filename_dict)
return train_test_data, y_pred_test, model_hdl
# --------------------------------------------
INPUT_MODEL = xgb.XGBClassifier()
MODEL = ModelHandler(INPUT_MODEL)
# hyperparams optimization
HYP_RANGES = {
# # defines the maximum depth of a single tree (regularization)
'max_depth': (5, 15),
# 'learning_rate': (0.01, 0.3), # learning rate
'n_estimators': (5, 10), # number of boosting trees
}
MODEL.optimize_params_bayes(DATA, HYP_RANGES, 'roc_auc')
# train and test the model with the updated hyperparameters
MODEL.train_test_model(DATA)
Y_PRED = MODEL.predict(DATA[2])
# Calculate the BDT efficiency as a function of the BDT score
EFFICIENCY, THRESHOLD = analysis_utils.bdt_efficiency_array(
DATA[3], Y_PRED, n_points=10)
# --------------------------------------------
# PLOTTING
# --------------------------------------------
FEATURES_DISTRIBUTIONS_PLOT = plot_utils.plot_distr(
[SIG_DF, BKG_DF], SIG_DF.columns)
CORRELATION_MATRIX_PLOT = plot_utils.plot_corr([SIG_DF, BKG_DF], SIG_DF.columns)
BDT_OUTPUT_PLOT = plot_utils.plot_output_train_test(MODEL, DATA)
ROC_CURVE_PLOT = plot_utils.plot_roc(DATA[3], Y_PRED)
PRECISION_RECALL_PLOT = plot_utils.plot_precision_recall(DATA[3], Y_PRED)
if TRAIN:
model_hdl.train_test_model(train_test_data)
model_file_name = str(f'models/{bin}_trained')
if OPTIMIZE:
model_file_name = str(
f'models/{bin}_optimized_trained')
model_hdl.dump_model_handler(model_file_name)
else:
if OPTIMIZED:
model_hdl.load_model_handler(
f'models/{bin}_optimized_trained')
else:
model_hdl.load_model_handler(f'models/{bin}_trained')
# get predictions for training and test set
test_y_score = model_hdl.predict(train_test_data[2])
train_y_score = model_hdl.predict(train_test_data[0])
# second condition needed because of issue with Qt libraries
if MAKE_TRAIN_TEST_PLOT and not MAKE_PRESELECTION_EFFICIENCY:
if not os.path.isdir(f'{PLOT_DIR}/train_test_out'):
os.mkdir(f'{PLOT_DIR}/train_test_out')
plot_utils.plot_output_train_test(model_hdl,
train_test_data,
logscale=True,
density=True,
labels=leg_labels)
plt.savefig(f'{PLOT_DIR}/train_test_out/{bin}_out')
plot_utils.plot_feature_imp(train_test_data[0],
train_test_data[1], model_hdl)
model_handler.set_training_columns(COLUMNS)
if OPTIMIZE:
model_handler.optimize_params_bayes(data,
HYPERPARAMS_RANGE,
'roc_auc',
init_points=10,
n_iter=10)
model_handler.train_test_model(data)
print("train test model")
print(
f'--- model trained and tested in {((time.time() - part_time) / 60):.2f} minutes ---\n'
)
y_pred = model_handler.predict(data[2])
data[2].insert(0, 'score', y_pred)
eff, tsd = analysis_utils.bdt_efficiency_array(
data[3], y_pred, n_points=1000)
score_from_eff_array = analysis_utils.score_from_efficiency_array(
data[3], y_pred, FIX_EFF_ARRAY)
fixed_eff_array = np.vstack(
(FIX_EFF_ARRAY, score_from_eff_array))
if SIGMA_MC:
ml_analysis.MC_sigma_array(data, fixed_eff_array,
cclass, ptbin, ctbin, split)
ml_analysis.save_ML_analysis(model_handler,
fixed_eff_array,
cent_class=cclass,
def get_skimmed_large_data(data_path,
cent_classes,
pt_bins,
ct_bins,
training_columns,
application_columns,
mode,
split=''):
print('\n++++++++++++++++++++++++++++++++++++++++++++++++++')
print('\nStarting BDT appplication on large data')
if mode == 3:
handlers_path = os.environ['HYPERML_MODELS_3'] + '/handlers'
efficiencies_path = os.environ['HYPERML_EFFICIENCIES_3']
if mode == 2:
handlers_path = os.environ['HYPERML_MODELS_2'] + '/handlers'
efficiencies_path = os.environ['HYPERML_EFFICIENCIES_2']
executor = ThreadPoolExecutor()
iterator = uproot.pandas.iterate(data_path,
'DataTable',
executor=executor,
reportfile=True)
df_applied = pd.DataFrame()
for current_file, data in iterator:
rename_df_columns(data)
print('current file: {}'.format(current_file))
print('start entry chunk: {}, stop entry chunk: {}'.format(
data.index[0], data.index[-1]))
for cclass in cent_classes:
for ptbin in zip(pt_bins[:-1], pt_bins[1:]):
for ctbin in zip(ct_bins[:-1], ct_bins[1:]):
info_string = '_{}{}_{}{}_{}{}'.format(
cclass[0], cclass[1], ptbin[0], ptbin[1], ctbin[0],
ctbin[1])
filename_handler = handlers_path + '/model_handler' + info_string + split + '.pkl'
filename_efficiencies = efficiencies_path + '/Eff_Score' + info_string + split + '.npy'
model_handler = ModelHandler()
model_handler.load_model_handler(filename_handler)
eff_score_array = np.load(filename_efficiencies)
tsd = eff_score_array[1][-1]
data_range = f'{ctbin[0]}<ct<{ctbin[1]} and {ptbin[0]}<pt<{ptbin[1]} and {cclass[0]}<=centrality<{cclass[1]}'
df_tmp = data.query(data_range)
df_tmp.insert(
0, 'score',
model_handler.predict(df_tmp[training_columns]))
df_tmp = df_tmp.query('score>@tsd')
df_tmp = df_tmp.loc[:, application_columns]
df_applied = df_applied.append(df_tmp,
ignore_index=True,
sort=False)
print(df_applied.info(memory_usage='deep'))
return df_applied
class Optimiserhipe4mltree:
# Class Attribute
species = "optimiser_hipe4mltree"
def __init__(self, data_param, binmin, binmax, training_var, bkg_sel,
hyper_pars):
self.logger = get_logger()
# directory
#self.do_mlprefilter = datap.get("doml_asprefilter", None)
self.dirmlout = data_param["ml"]["mlout"]
self.dirmlplot = data_param["ml"]["mlplot"]
#if self.do_mlprefilter is True:
# self.dirmodel = self.dirmodel + "/prefilter"
# self.dirmlplot = self.dirmlplot + "/prefilter"
#if self.do_mlprefilter is False:
# self.dirmodel = self.dirmodel + "/analysis"
# self.dirmlplot = self.dirmlplot + "/analysis"
self.inputtreedata = "/Users/lvermunt/cernbox/Analyses/ML/input/hipe4mlTTree/data.root"
self.inputtreemc = "/Users/lvermunt/cernbox/Analyses/ML/input/hipe4mlTTree/prompt.root"
self.v_train = None
self.p_binmin = binmin
self.p_binmax = binmax
self.s_selsigml = ""
self.s_selbkgml = bkg_sel #"inv_mass < 1.82 or 1.92 < inv_mass < 2.00"
self.v_bkgoversigfrac = 3
self.v_sig = 1
self.v_bkg = 0
self.rnd_splt = data_param["ml"]["rnd_splt"]
self.test_frac = data_param["ml"]["test_frac"]
self.prompthandler = None
self.datahandler = None
self.bkghandler = None
self.traintestdata = None
self.ypredtrain_hipe4ml = None
self.ypredtest_hipe4ml = None
self.preparesample()
self.p_hipe4ml_model = None
self.v_hipe4ml_pars = hyper_pars
self.load_hipe4mlmodel()
self.bayesoptconfig_hipe4ml = data_param["hipe4ml"]["hyper_par_opt"][
"bayes_opt_config"]
self.average_method_hipe4ml = data_param["hipe4ml"]["roc_auc_average"]
self.nfold_hipe4ml = data_param["hipe4ml"]["hyper_par_opt"]["nfolds"]
self.init_points = data_param["hipe4ml"]["hyper_par_opt"]["initpoints"]
self.n_iter_hipe4ml = data_param["hipe4ml"]["hyper_par_opt"]["niter"]
self.njobs_hipe4ml = data_param["hipe4ml"]["hyper_par_opt"]["njobs"]
self.roc_method_hipe4ml = data_param["hipe4ml"]["roc_auc_approach"]
self.raw_output_hipe4ml = data_param["hipe4ml"]["raw_output"]
self.train_test_log_hipe4ml = data_param["hipe4ml"]["train_test_log"]
self.multiclass_labels = data_param["ml"].get("multiclass_labels",
None)
self.logger.info("Using the following training variables: %s",
self.v_train)
def preparesample(self):
self.logger.info("Prepare Sample for hipe4ml")
self.signalhandler = TreeHandler(self.inputtreemc, 'treeMLDplus')
nsigcand = self.signalhandler.get_n_cand()
self.datahandler = TreeHandler(self.inputtreedata, 'treeMLDplus')
self.bkghandler = self.datahandler.get_subset(self.s_selbkgml,
size=nsigcand *
self.v_bkgoversigfrac)
self.traintestdata = train_test_generator(
[self.signalhandler, self.bkghandler], [self.v_sig, self.v_bkg],
test_size=self.test_frac,
random_state=self.rnd_splt)
def load_hipe4mlmodel(self):
self.logger.info("Loading hipe4ml model")
self.v_train = self.signalhandler.get_var_names()
self.v_train.remove('inv_mass')
self.v_train.remove('pt_cand')
model_xgboost = xgb.XGBClassifier()
self.p_hipe4ml_model = ModelHandler(model_xgboost, self.v_train)
def set_hipe4ml_modelpar(self):
self.logger.info("Setting hipe4ml hyperparameters")
self.p_hipe4ml_model.set_model_params(self.v_hipe4ml_pars)
def do_hipe4mlhyperparopti(self):
self.logger.info("Optimising hipe4ml hyperparameters (Bayesian)")
if not (self.average_method_hipe4ml in ['macro', 'weighted']
and self.roc_method_hipe4ml in ['ovo', 'ovr']):
self.logger.fatal("Selected ROC configuration is not valid!")
if self.average_method_hipe4ml == 'weighted':
metric = f'roc_auc_{self.roc_method_hipe4ml}_{self.average_method_hipe4ml}'
else:
metric = f'roc_auc_{self.roc_method_hipe4ml}'
hypparsfile = f'{self.dirmlout}/HyperParOpt_pT_{self.p_binmin}_{self.p_binmax}.txt'
outfilehyppars = open(hypparsfile, 'wt')
sys.stdout = outfilehyppars
self.p_hipe4ml_model.optimize_params_bayes(self.traintestdata,
self.bayesoptconfig_hipe4ml,
metric, self.nfold_hipe4ml,
self.init_points,
self.n_iter_hipe4ml,
self.njobs_hipe4ml)
outfilehyppars.close()
sys.stdout = sys.__stdout__
self.logger.info("Performing hyper-parameters optimisation: Done!")
def do_hipe4mltrain(self):
self.logger.info("Training + testing hipe4ml model")
t0 = time.time()
self.p_hipe4ml_model.train_test_model(self.traintestdata,
self.average_method_hipe4ml,
self.roc_method_hipe4ml)
self.ypredtrain_hipe4ml = self.p_hipe4ml_model.predict(
self.traintestdata[0], self.raw_output_hipe4ml)
self.ypredtest_hipe4ml = self.p_hipe4ml_model.predict(
self.traintestdata[2], self.raw_output_hipe4ml)
modelhandlerfile = f'{self.dirmlout}/ModelHandler_pT_{self.p_binmin}_{self.p_binmax}.pkl'
self.p_hipe4ml_model.dump_model_handler(modelhandlerfile)
modelfile = f'{self.dirmlout}/ModelHandler_pT_{self.p_binmin}_{self.p_binmax}.model'
self.p_hipe4ml_model.dump_original_model(modelfile)
self.logger.info("Training + testing hipe4ml: Done!")
self.logger.info("Time elapsed = %.3f", time.time() - t0)
def do_hipe4mlplot(self):
self.logger.info("Plotting hipe4ml model")
leglabels = ["Background", "Prompt signal"]
outputlabels = ["Bkg", "SigPrompt"]
# _____________________________________________
plot_utils.plot_distr([self.bkghandler, self.signalhandler],
self.v_train, 100, leglabels)
plt.subplots_adjust(left=0.06,
bottom=0.06,
right=0.99,
top=0.96,
hspace=0.55,
wspace=0.55)
figname = f'{self.dirmlplot}/DistributionsAll_pT_{self.p_binmin}_{self.p_binmax}.pdf'
plt.savefig(figname)
plt.close('all')
# _____________________________________________
corrmatrixfig = plot_utils.plot_corr(
[self.bkghandler, self.signalhandler], self.v_train, leglabels)
for figg, labb in zip(corrmatrixfig, outputlabels):
plt.figure(figg.number)
plt.subplots_adjust(left=0.2, bottom=0.25, right=0.95, top=0.9)
figname = f'{self.dirmlplot}/CorrMatrix{labb}_pT_{self.p_binmin}_{self.p_binmax}.pdf'
figg.savefig(figname)
# _____________________________________________
plt.rcParams["figure.figsize"] = (10, 7)
mloutputfig = plot_utils.plot_output_train_test(
self.p_hipe4ml_model,
self.traintestdata,
80,
self.raw_output_hipe4ml,
leglabels,
self.train_test_log_hipe4ml,
density=True)
figname = f'{self.dirmlplot}/MLOutputDistr_pT_{self.p_binmin}_{self.p_binmax}.pdf'
mloutputfig.savefig(figname)
# _____________________________________________
plt.rcParams["figure.figsize"] = (10, 9)
roccurvefig = plot_utils.plot_roc(self.traintestdata[3],
self.ypredtest_hipe4ml, None,
leglabels,
self.average_method_hipe4ml,
self.roc_method_hipe4ml)
figname = f'{self.dirmlplot}/ROCCurveAll_pT_{self.p_binmin}_{self.p_binmax}.pdf'
roccurvefig.savefig(figname)
# _____________________________________________
plt.rcParams["figure.figsize"] = (10, 9)
roccurvettfig = plot_utils.plot_roc_train_test(
self.traintestdata[3], self.ypredtest_hipe4ml,
self.traintestdata[1], self.ypredtrain_hipe4ml, None, leglabels,
self.average_method_hipe4ml, self.roc_method_hipe4ml)
figname = f'{self.dirmlplot}/ROCCurveTrainTest_pT_{self.p_binmin}_{self.p_binmax}.pdf'
roccurvettfig.savefig(figname)
# _____________________________________________
precisionrecallfig = plot_utils.plot_precision_recall(
self.traintestdata[3], self.ypredtest_hipe4ml, leglabels)
figname = f'{self.dirmlplot}/PrecisionRecallAll_pT_{self.p_binmin}_{self.p_binmax}.pdf'
precisionrecallfig.savefig(figname)
# _____________________________________________
plt.rcParams["figure.figsize"] = (12, 7)
featuresimportancefig = plot_utils.plot_feature_imp(
self.traintestdata[2][self.v_train], self.traintestdata[3],
self.p_hipe4ml_model, leglabels)
for i in range(0, len(featuresimportancefig)):
figname = (f'{self.dirmlplot}/FeatureImportanceOpt{i}_'
f'pT_{self.p_binmin}_{self.p_binmax}.pdf')
featuresimportancefig[i].savefig(figname)