def test_manual_pipeline(sampled_app_train_test, sampled_app_roles,
binary_task):
train, test = sampled_app_train_test
pd_dataset = PandasDataset(train,
roles_parser(sampled_app_roles),
task=binary_task)
selector_iterator = FoldsIterator(pd_dataset, 1)
pipe = LGBSimpleFeatures()
model0 = BoostLGBM(default_params={
"learning_rate": 0.05,
"num_leaves": 64,
"seed": 0,
"num_threads": 5,
})
mbie = ModelBasedImportanceEstimator()
selector = ImportanceCutoffSelector(pipe, model0, mbie, cutoff=10)
selector.fit(selector_iterator)
pipe = LGBSimpleFeatures()
params_tuner1 = OptunaTuner(n_trials=10, timeout=300)
model1 = BoostLGBM(default_params={
"learning_rate": 0.05,
"num_leaves": 128
})
params_tuner2 = OptunaTuner(n_trials=100, timeout=300)
model2 = BoostLGBM(default_params={
"learning_rate": 0.025,
"num_leaves": 64
})
total = MLPipeline(
[(model1, params_tuner1), (model2, params_tuner2)],
pre_selection=selector,
features_pipeline=pipe,
post_selection=None,
)
train_valid = FoldsIterator(pd_dataset)
total.fit_predict(train_valid)
total.predict(pd_dataset)
with open("automl.pickle", "wb") as f:
pickle.dump(total, f)
with open("automl.pickle", "rb") as f:
total = pickle.load(f)
total.predict(pd_dataset)
os.remove("automl.pickle")
data["allnan"] = np.nan
data.drop(["DAYS_BIRTH", "DAYS_EMPLOYED"], axis=1, inplace=True)
data["TARGET"] = data["TARGET"]
print("Features modification finished")
print("Split data...")
train, test = train_test_split(data, test_size=0.2, random_state=42)
train.reset_index(drop=True, inplace=True)
test.reset_index(drop=True, inplace=True)
print("Data splitted. Parts sizes: train_data = {}, test_data = {}".format(
train.shape, test.shape))
print("Start creation selector_0...")
feat_sel_0 = LGBSimpleFeatures()
mod_sel_0 = BoostLGBM()
imp_sel_0 = ModelBasedImportanceEstimator()
selector_0 = ImportanceCutoffSelector(feat_sel_0,
mod_sel_0,
imp_sel_0,
cutoff=0)
print("End creation selector_0...")
print("Start creation gbm_0...")
feats_gbm_0 = LGBAdvancedPipeline()
gbm_0 = BoostLGBM()
gbm_1 = BoostLGBM()
tuner_0 = OptunaTuner(n_trials=100, timeout=30, fit_on_holdout=True)
gbm_lvl0 = MLPipeline(
[(gbm_0, tuner_0), gbm_1],
def test_permutation_importance_based_iterative_selector():
logging.basicConfig(format="[%(asctime)s] (%(levelname)s): %(message)s",
level=logging.DEBUG)
logging.debug("Load data...")
data = pd.read_csv("./examples/data/sampled_app_train.csv")
logging.debug("Data loaded")
logging.debug("Features modification from user side...")
data["BIRTH_DATE"] = (
np.datetime64("2018-01-01") +
data["DAYS_BIRTH"].astype(np.dtype("timedelta64[D]"))).astype(str)
data["EMP_DATE"] = (np.datetime64("2018-01-01") +
np.clip(data["DAYS_EMPLOYED"], None, 0).astype(
np.dtype("timedelta64[D]"))).astype(str)
data["constant"] = 1
data["allnan"] = np.nan
data.drop(["DAYS_BIRTH", "DAYS_EMPLOYED"], axis=1, inplace=True)
logging.debug("Features modification finished")
logging.debug("Split data...")
train_data, test_data = train_test_split(data,
test_size=2000,
stratify=data["TARGET"],
random_state=13)
train_data.reset_index(drop=True, inplace=True)
test_data.reset_index(drop=True, inplace=True)
logging.debug(
"Data splitted. Parts sizes: train_data = {}, test_data = {}".format(
train_data.shape, test_data.shape))
logging.debug("Create task...")
task = Task("binary")
logging.debug("Task created")
logging.debug("Create reader...")
reader = PandasToPandasReader(task, cv=5, random_state=1)
logging.debug("Reader created")
# selector parts
logging.debug("Create feature selector")
model0 = BoostLGBM(default_params={
"learning_rate": 0.05,
"num_leaves": 64,
"seed": 42,
"num_threads": 5,
})
pipe0 = LGBSimpleFeatures()
pie = NpPermutationImportanceEstimator()
selector = NpIterativeFeatureSelector(pipe0,
model0,
pie,
feature_group_size=1,
max_features_cnt_in_result=15)
logging.debug("Feature selector created")
# pipeline 1 level parts
logging.debug("Start creation pipeline_1...")
pipe = LGBSimpleFeatures()
logging.debug("\t ParamsTuner1 and Model1...")
model1 = BoostLGBM(default_params={
"learning_rate": 0.05,
"num_leaves": 128,
"seed": 1,
"num_threads": 5,
})
logging.debug("\t Tuner1 and model1 created")
logging.debug("\t ParamsTuner2 and Model2...")
params_tuner2 = OptunaTuner(n_trials=100, timeout=100)
model2 = BoostLGBM(default_params={
"learning_rate": 0.025,
"num_leaves": 64,
"seed": 2,
"num_threads": 5,
})
logging.debug("\t Tuner2 and model2 created")
logging.debug("\t Pipeline1...")
pipeline_lvl1 = MLPipeline(
[model1, (model2, params_tuner2)],
pre_selection=selector,
features_pipeline=pipe,
post_selection=None,
)
logging.debug("Pipeline1 created")
# pipeline 2 level parts
logging.debug("Start creation pipeline_2...")
pipe1 = LGBSimpleFeatures()
logging.debug("\t ParamsTuner and Model...")
model = BoostLGBM(
default_params={
"learning_rate": 0.05,
"num_leaves": 64,
"max_bin": 1024,
"seed": 3,
"num_threads": 5,
},
freeze_defaults=True,
)
logging.debug("\t Tuner and model created")
logging.debug("\t Pipeline2...")
pipeline_lvl2 = MLPipeline([model],
pre_selection=None,
features_pipeline=pipe1,
post_selection=None)
logging.debug("Pipeline2 created")
logging.debug("Create AutoML pipeline...")
automl = AutoML(
reader,
[
[pipeline_lvl1],
[pipeline_lvl2],
],
skip_conn=False,
)
logging.debug("AutoML pipeline created...")
logging.debug("Start AutoML pipeline fit_predict...")
start_time = time.time()
oof_pred = automl.fit_predict(train_data, roles={"target": "TARGET"})
logging.debug(
"AutoML pipeline fitted and predicted. Time = {:.3f} sec".format(
time.time() - start_time))
logging.debug("Feature importances of selector:\n{}".format(
selector.get_features_score()))
logging.debug("oof_pred:\n{}\nShape = {}".format(oof_pred, oof_pred.shape))
logging.debug("Feature importances of top level algorithm:\n{}".format(
automl.levels[-1][0].ml_algos[0].get_features_score()))
logging.debug(
"Feature importances of lowest level algorithm - model 0:\n{}".format(
automl.levels[0][0].ml_algos[0].get_features_score()))
logging.debug(
"Feature importances of lowest level algorithm - model 1:\n{}".format(
automl.levels[0][0].ml_algos[1].get_features_score()))
test_pred = automl.predict(test_data)
logging.debug("Prediction for test data:\n{}\nShape = {}".format(
test_pred, test_pred.shape))
logging.debug("Check scores...")
logging.debug("OOF score: {}".format(
roc_auc_score(train_data["TARGET"].values, oof_pred.data[:, 0])))
logging.debug("TEST score: {}".format(
roc_auc_score(test_data["TARGET"].values, test_pred.data[:, 0])))
logging.debug("Pickle automl")
with open("automl.pickle", "wb") as f:
pickle.dump(automl, f)
logging.debug("Load pickled automl")
with open("automl.pickle", "rb") as f:
automl = pickle.load(f)
logging.debug("Predict loaded automl")
test_pred = automl.predict(test_data)
logging.debug("TEST score, loaded: {}".format(
roc_auc_score(test_data["TARGET"].values, test_pred.data[:, 0])))
os.remove("automl.pickle")
},
},
],
["TARGET", "TARGET", "AMT_CREDIT", "AMT_CREDIT", "AMT_CREDIT"],
):
print("Create task..")
task = Task(**task_params)
print("Task created")
print("Create reader...")
reader = PandasToPandasReader(task, cv=5, random_state=1)
print("Reader created")
# pipeline 1 level parts
print("Start creation pipeline_1...")
pipe = LGBSimpleFeatures()
print("\t ParamsTuner2 and Model2...")
model2 = BoostLGBM(default_params={
"learning_rate": 0.025,
"num_leaves": 64,
"seed": 2,
"num_threads": 5,
})
print("\t Tuner2 and model2 created")
print("\t Pipeline1...")
pipeline_lvl1 = MLPipeline(
[model2],
pre_selection=None, # selector,
features_pipeline=pipe,
def test_permutation_importance_based_iterative_selector():
logging.basicConfig(format='[%(asctime)s] (%(levelname)s): %(message)s',
level=logging.DEBUG)
logging.debug('Load data...')
data = pd.read_csv('../example_data/test_data_files/sampled_app_train.csv')
logging.debug('Data loaded')
logging.debug('Features modification from user side...')
data['BIRTH_DATE'] = (
np.datetime64('2018-01-01') +
data['DAYS_BIRTH'].astype(np.dtype('timedelta64[D]'))).astype(str)
data['EMP_DATE'] = (np.datetime64('2018-01-01') +
np.clip(data['DAYS_EMPLOYED'], None, 0).astype(
np.dtype('timedelta64[D]'))).astype(str)
data['constant'] = 1
data['allnan'] = np.nan
data.drop(['DAYS_BIRTH', 'DAYS_EMPLOYED'], axis=1, inplace=True)
logging.debug('Features modification finished')
logging.debug('Split data...')
train_data, test_data = train_test_split(data,
test_size=2000,
stratify=data['TARGET'],
random_state=13)
train_data.reset_index(drop=True, inplace=True)
test_data.reset_index(drop=True, inplace=True)
logging.debug(
'Data splitted. Parts sizes: train_data = {}, test_data = {}'.format(
train_data.shape, test_data.shape))
logging.debug('Create task...')
task = Task('binary')
logging.debug('Task created')
logging.debug('Create reader...')
reader = PandasToPandasReader(task, cv=5, random_state=1)
logging.debug('Reader created')
# selector parts
logging.debug('Create feature selector')
model0 = BoostLGBM(default_params={
'learning_rate': 0.05,
'num_leaves': 64,
'seed': 42,
'num_threads': 5
})
pipe0 = LGBSimpleFeatures()
pie = NpPermutationImportanceEstimator()
selector = NpIterativeFeatureSelector(pipe0,
model0,
pie,
feature_group_size=1,
max_features_cnt_in_result=15)
logging.debug('Feature selector created')
# pipeline 1 level parts
logging.debug('Start creation pipeline_1...')
pipe = LGBSimpleFeatures()
logging.debug('\t ParamsTuner1 and Model1...')
model1 = BoostLGBM(default_params={
'learning_rate': 0.05,
'num_leaves': 128,
'seed': 1,
'num_threads': 5
})
logging.debug('\t Tuner1 and model1 created')
logging.debug('\t ParamsTuner2 and Model2...')
params_tuner2 = OptunaTuner(n_trials=100, timeout=100)
model2 = BoostLGBM(default_params={
'learning_rate': 0.025,
'num_leaves': 64,
'seed': 2,
'num_threads': 5
})
logging.debug('\t Tuner2 and model2 created')
logging.debug('\t Pipeline1...')
pipeline_lvl1 = MLPipeline([model1, (model2, params_tuner2)],
pre_selection=selector,
features_pipeline=pipe,
post_selection=None)
logging.debug('Pipeline1 created')
# pipeline 2 level parts
logging.debug('Start creation pipeline_2...')
pipe1 = LGBSimpleFeatures()
logging.debug('\t ParamsTuner and Model...')
model = BoostLGBM(default_params={
'learning_rate': 0.05,
'num_leaves': 64,
'max_bin': 1024,
'seed': 3,
'num_threads': 5
},
freeze_defaults=True)
logging.debug('\t Tuner and model created')
logging.debug('\t Pipeline2...')
pipeline_lvl2 = MLPipeline([model],
pre_selection=None,
features_pipeline=pipe1,
post_selection=None)
logging.debug('Pipeline2 created')
logging.debug('Create AutoML pipeline...')
automl = AutoML(reader, [
[pipeline_lvl1],
[pipeline_lvl2],
],
skip_conn=False)
logging.debug('AutoML pipeline created...')
logging.debug('Start AutoML pipeline fit_predict...')
start_time = time.time()
oof_pred = automl.fit_predict(train_data, roles={'target': 'TARGET'})
logging.debug(
'AutoML pipeline fitted and predicted. Time = {:.3f} sec'.format(
time.time() - start_time))
logging.debug('Feature importances of selector:\n{}'.format(
selector.get_features_score()))
logging.debug('oof_pred:\n{}\nShape = {}'.format(oof_pred, oof_pred.shape))
logging.debug('Feature importances of top level algorithm:\n{}'.format(
automl.levels[-1][0].ml_algos[0].get_features_score()))
logging.debug(
'Feature importances of lowest level algorithm - model 0:\n{}'.format(
automl.levels[0][0].ml_algos[0].get_features_score()))
logging.debug(
'Feature importances of lowest level algorithm - model 1:\n{}'.format(
automl.levels[0][0].ml_algos[1].get_features_score()))
test_pred = automl.predict(test_data)
logging.debug('Prediction for test data:\n{}\nShape = {}'.format(
test_pred, test_pred.shape))
logging.debug('Check scores...')
logging.debug('OOF score: {}'.format(
roc_auc_score(train_data['TARGET'].values, oof_pred.data[:, 0])))
logging.debug('TEST score: {}'.format(
roc_auc_score(test_data['TARGET'].values, test_pred.data[:, 0])))
logging.debug('Pickle automl')
with open('automl.pickle', 'wb') as f:
pickle.dump(automl, f)
logging.debug('Load pickled automl')
with open('automl.pickle', 'rb') as f:
automl = pickle.load(f)
logging.debug('Predict loaded automl')
test_pred = automl.predict(test_data)
logging.debug('TEST score, loaded: {}'.format(
roc_auc_score(test_data['TARGET'].values, test_pred.data[:, 0])))
os.remove('automl.pickle')
def test_manual_pipeline():
# Read data from file
logging.debug('Read data from file')
data = pd.read_csv('../example_data/test_data_files/sampled_app_train.csv',
usecols=[
'TARGET', 'NAME_CONTRACT_TYPE', 'AMT_CREDIT',
'NAME_TYPE_SUITE', 'AMT_GOODS_PRICE', 'DAYS_BIRTH',
'DAYS_EMPLOYED'
])
# Fix dates and convert to date type
logging.debug('Fix dates and convert to date type')
data['BIRTH_DATE'] = np.datetime64(
'2018-01-01') + data['DAYS_BIRTH'].astype(np.dtype('timedelta64[D]'))
data['EMP_DATE'] = np.datetime64('2018-01-01') + np.clip(
data['DAYS_EMPLOYED'], None, 0).astype(np.dtype('timedelta64[D]'))
data.drop(['DAYS_BIRTH', 'DAYS_EMPLOYED'], axis=1, inplace=True)
# Create folds
logging.debug('Create folds')
data['__fold__'] = np.random.randint(0, 5, len(data))
# Print data head
logging.debug('Print data head')
print(data.head())
# # Set roles for columns
logging.debug('Set roles for columns')
check_roles = {
TargetRole(): 'TARGET',
CategoryRole(dtype=str): ['NAME_CONTRACT_TYPE', 'NAME_TYPE_SUITE'],
NumericRole(np.float32): ['AMT_CREDIT', 'AMT_GOODS_PRICE'],
DatetimeRole(seasonality=['y', 'm', 'wd']): ['BIRTH_DATE', 'EMP_DATE'],
FoldsRole(): '__fold__'
}
# create Task
task = Task('binary')
# # Creating PandasDataSet
logging.debug('Creating PandasDataset')
start_time = time.time()
pd_dataset = PandasDataset(data, roles_parser(check_roles), task=task)
logging.debug(
'PandasDataset created. Time = {:.3f} sec'.format(time.time() -
start_time))
# # Print pandas dataset feature roles
logging.debug('Print pandas dataset feature roles')
roles = pd_dataset.roles
for role in roles:
logging.debug('{}: {}'.format(role, roles[role]))
# # Feature selection part
logging.debug('Feature selection part')
selector_iterator = FoldsIterator(pd_dataset, 1)
logging.debug('Selection iterator created')
model = BoostLGBM()
pipe = LGBSimpleFeatures()
logging.debug('Pipe and model created')
model0 = BoostLGBM(default_params={
'learning_rate': 0.05,
'num_leaves': 64,
'seed': 0,
'num_threads': 5
})
mbie = ModelBasedImportanceEstimator()
selector = ImportanceCutoffSelector(pipe, model0, mbie, cutoff=10)
start_time = time.time()
selector.fit(selector_iterator)
logging.debug(
'Feature selector fitted. Time = {:.3f} sec'.format(time.time() -
start_time))
logging.debug('Feature selector scores:')
logging.debug('\n{}'.format(selector.get_features_score()))
# # Build AutoML pipeline
logging.debug('Start building AutoML pipeline')
pipe = LGBSimpleFeatures()
logging.debug('Pipe created')
params_tuner1 = OptunaTuner(n_trials=10, timeout=300)
model1 = BoostLGBM(default_params={
'learning_rate': 0.05,
'num_leaves': 128
})
logging.debug('Tuner1 and model1 created')
params_tuner2 = OptunaTuner(n_trials=100, timeout=300)
model2 = BoostLGBM(default_params={
'learning_rate': 0.025,
'num_leaves': 64
})
logging.debug('Tuner2 and model2 created')
total = MLPipeline([(model1, params_tuner1), (model2, params_tuner2)],
pre_selection=selector,
features_pipeline=pipe,
post_selection=None)
logging.debug('Finished building AutoML pipeline')
# # Create full train iterator
logging.debug('Full train valid iterator creation')
train_valid = FoldsIterator(pd_dataset)
logging.debug('Full train valid iterator created')
# # Fit predict using pipeline
logging.debug('Start AutoML pipeline fit_predict')
start_time = time.time()
pred = total.fit_predict(train_valid)
logging.debug(
'Fit_predict finished. Time = {:.3f} sec'.format(time.time() -
start_time))
# # Check preds
logging.debug('Preds:')
logging.debug('\n{}'.format(pred))
logging.debug('Preds.shape = {}'.format(pred.shape))
# # Predict full train dataset
logging.debug('Predict full train dataset')
start_time = time.time()
train_pred = total.predict(pd_dataset)
logging.debug('Predict finished. Time = {:.3f} sec'.format(time.time() -
start_time))
logging.debug('Preds:')
logging.debug('\n{}'.format(train_pred))
logging.debug('Preds.shape = {}'.format(train_pred.shape))
logging.debug('Pickle automl')
with open('automl.pickle', 'wb') as f:
pickle.dump(total, f)
logging.debug('Load pickled automl')
with open('automl.pickle', 'rb') as f:
total = pickle.load(f)
logging.debug('Predict loaded automl')
train_pred = total.predict(pd_dataset)
os.remove('automl.pickle')
# # Check preds feature names
logging.debug('Preds features: {}'.format(train_pred.features))
# # Check model feature scores
logging.debug('Feature scores for model_1:\n{}'.format(
model1.get_features_score()))
logging.debug('Feature scores for model_2:\n{}'.format(
model2.get_features_score()))
def test_boostlgbm_and_linearlbfgs_in_one_automl_pipeline():
np.random.seed(42)
logging.basicConfig(format='[%(asctime)s] (%(levelname)s): %(message)s', level=logging.DEBUG)
logging.debug('Load data...')
data = pd.read_csv('../example_data/test_data_files/sampled_app_train.csv')
logging.debug('Data loaded')
logging.debug('Features modification from user side...')
data['BIRTH_DATE'] = (np.datetime64('2018-01-01') + data['DAYS_BIRTH'].astype(np.dtype('timedelta64[D]'))).astype(str)
data['EMP_DATE'] = (np.datetime64('2018-01-01') + np.clip(data['DAYS_EMPLOYED'], None, 0).astype(np.dtype('timedelta64[D]'))
).astype(str)
data['report_dt'] = np.datetime64('2018-01-01')
data['constant'] = 1
data['allnan'] = np.nan
data.drop(['DAYS_BIRTH', 'DAYS_EMPLOYED'], axis=1, inplace=True)
data['TARGET'] = data['TARGET']
logging.debug('Features modification finished')
logging.debug('Split data...')
train, test = train_test_split(data, test_size=0.2, random_state=42)
train.reset_index(drop=True, inplace=True)
test.reset_index(drop=True, inplace=True)
logging.debug('Data splitted. Parts sizes: train_data = {}, test_data = {}'
.format(train.shape, test.shape))
logging.debug('Start creation selector_0...')
feat_sel_0 = LGBSimpleFeatures()
mod_sel_0 = BoostLGBM()
imp_sel_0 = ModelBasedImportanceEstimator()
selector_0 = ImportanceCutoffSelector(feat_sel_0, mod_sel_0, imp_sel_0, cutoff=0)
logging.debug('End creation selector_0...')
logging.debug('Start creation gbm_0...')
feats_gbm_0 = LGBAdvancedPipeline()
gbm_0 = BoostLGBM()
gbm_1 = BoostLGBM()
tuner_0 = OptunaTuner(n_trials=100, timeout=30, fit_on_holdout=True)
gbm_lvl0 = MLPipeline([
(gbm_0, tuner_0),
gbm_1
],
pre_selection=selector_0,
features_pipeline=feats_gbm_0, post_selection=None)
logging.debug('End creation gbm_0...')
logging.debug('Start creation reg_0...')
feats_reg_0 = LinearFeatures(output_categories=True)
reg_0 = LinearLBFGS()
reg_lvl0 = MLPipeline([
reg_0
],
pre_selection=None,
features_pipeline=feats_reg_0,
post_selection=HighCorrRemoval(corr_co=1))
logging.debug('End creation reg_0...')
logging.debug('Start creation composed selector...')
feat_sel_1 = LGBSimpleFeatures()
mod_sel_1 = BoostLGBM()
imp_sel_1 = NpPermutationImportanceEstimator()
selector_1 = NpIterativeFeatureSelector(feat_sel_1, mod_sel_1, imp_sel_1, feature_group_size=1)
logging.debug('End creation composed selector...')
logging.debug('Start creation reg_l1_0...')
feats_reg_1 = LinearFeatures(output_categories=False)
reg_1 = LinearL1CD()
reg_l1_lvl0 = MLPipeline([
reg_1
],
pre_selection=selector_1,
features_pipeline=feats_reg_1,
post_selection=HighCorrRemoval())
logging.debug('End creation reg_l1_0...')
logging.debug('Start creation blending...')
feats_reg_2 = LinearFeatures(output_categories=True)
reg_2 = LinearLBFGS()
reg_lvl1 = MLPipeline([
reg_2
],
pre_selection=None,
features_pipeline=feats_reg_2,
post_selection=HighCorrRemoval(corr_co=1))
logging.debug('End creation blending...')
logging.debug('Start creation automl...')
reader = PandasToPandasReader(Task('binary', ), samples=None, max_nan_rate=1, max_constant_rate=1)
automl = AutoML(reader, [
[gbm_lvl0, reg_lvl0, reg_l1_lvl0],
[reg_lvl1],
], skip_conn=False, blender=MeanBlender())
logging.debug('End creation automl...')
logging.debug('Start fit automl...')
roles = {'target': 'TARGET',
DatetimeRole(base_date=True, seasonality=(), base_feats=False): 'report_dt',
}
oof_pred = automl.fit_predict(train, roles=roles)
logging.debug('End fit automl...')
test_pred = automl.predict(test)
logging.debug('Prediction for test data:\n{}\nShape = {}'
.format(test_pred, test_pred.shape))
not_nan = np.any(~np.isnan(oof_pred.data), axis=1)
logging.debug('Check scores...')
print('OOF score: {}'.format(roc_auc_score(train[roles['target']].values[not_nan], oof_pred.data[not_nan][:, 0])))
print('TEST score: {}'.format(roc_auc_score(test[roles['target']].values, test_pred.data[:, 0])))
logging.debug('Pickle automl')
with open('automl.pickle', 'wb') as f:
pickle.dump(automl, f)
logging.debug('Load pickled automl')
with open('automl.pickle', 'rb') as f:
automl = pickle.load(f)
logging.debug('Predict loaded automl')
test_pred = automl.predict(test)
logging.debug('TEST score, loaded: {}'.format(roc_auc_score(test['TARGET'].values, test_pred.data[:, 0])))
os.remove('automl.pickle')
# selector parts
print("Create feature selector")
model01 = BoostLGBM(default_params={
"learning_rate": 0.05,
"num_leaves": 64,
"seed": 42,
"num_threads": 5,
})
model02 = BoostLGBM(default_params={
"learning_rate": 0.05,
"num_leaves": 64,
"seed": 42,
"num_threads": 5,
})
pipe0 = LGBSimpleFeatures()
pie = NpPermutationImportanceEstimator()
pie1 = ModelBasedImportanceEstimator()
sel1 = ImportanceCutoffSelector(pipe0, model01, pie1, cutoff=0)
sel2 = NpIterativeFeatureSelector(pipe0,
model02,
pie,
feature_group_size=1,
max_features_cnt_in_result=15)
selector = ComposedSelector([sel1, sel2])
print("Feature selector created")
# pipeline 1 level parts
print("Start creation pipeline_1...")
pipe = LGBSimpleFeatures()
def test_multiclass_task_with_catboost():
data = pd.read_csv('../example_data/test_data_files/sampled_app_train.csv')
data['BIRTH_DATE'] = (
np.datetime64('2018-01-01') +
data['DAYS_BIRTH'].astype(np.dtype('timedelta64[D]'))).astype(str)
data['EMP_DATE'] = (np.datetime64('2018-01-01') +
np.clip(data['DAYS_EMPLOYED'], None, 0).astype(
np.dtype('timedelta64[D]'))).astype(str)
data['report_dt'] = np.datetime64('2018-01-01')
data['constant'] = 1
data['allnan'] = np.nan
data.drop(['DAYS_BIRTH', 'DAYS_EMPLOYED'], axis=1, inplace=True)
data['TARGET'] = np.where(
np.random.rand(data.shape[0]) > .5, 2, data['TARGET'].values)
train, test = train_test_split(data, test_size=2000, random_state=42)
# ======================================================================================
logging.debug('Create timer...')
timer = PipelineTimer(600, mode=2)
logging.debug('Timer created...')
# ======================================================================================
logging.debug('Create selector...')
timer_gbm = timer.get_task_timer('gbm')
feat_sel_0 = LGBSimpleFeatures()
mod_sel_0 = BoostCB(timer=timer_gbm)
imp_sel_0 = ModelBasedImportanceEstimator()
selector_0 = ImportanceCutoffSelector(
feat_sel_0,
mod_sel_0,
imp_sel_0,
cutoff=0,
)
logging.debug('Selector created...')
# ======================================================================================
logging.debug('Create gbms...')
feats_gbm_0 = LGBAdvancedPipeline(top_intersections=4, feats_imp=imp_sel_0)
timer_gbm_0 = timer.get_task_timer('gbm')
timer_gbm_1 = timer.get_task_timer('gbm')
gbm_0 = BoostCB(timer=timer_gbm_0, default_params={"devices": "0"})
gbm_1 = BoostCB(timer=timer_gbm_1, default_params={"devices": "0"})
tuner_0 = OptunaTuner(n_trials=10, timeout=10, fit_on_holdout=True)
gbm_lvl0 = MLPipeline([(gbm_0, tuner_0), gbm_1],
pre_selection=selector_0,
features_pipeline=feats_gbm_0,
post_selection=None)
logging.debug('Gbms created...')
# ======================================================================================
logging.debug('Create linear...')
feats_reg_0 = LinearFeatures(output_categories=True, sparse_ohe='auto')
timer_reg = timer.get_task_timer('reg')
reg_0 = LinearLBFGS(timer=timer_reg)
reg_lvl0 = MLPipeline([reg_0],
pre_selection=None,
features_pipeline=feats_reg_0,
post_selection=None)
logging.debug('Linear created...')
# ======================================================================================
logging.debug('Create reader...')
reader = PandasToPandasReader(Task(
'multiclass',
metric='crossentropy',
),
samples=None,
max_nan_rate=1,
max_constant_rate=1,
advanced_roles=True,
drop_score_co=-1,
n_jobs=1)
logging.debug('Reader created...')
# ======================================================================================
logging.debug('Create blender...')
blender = WeightedBlender()
logging.debug('Blender created...')
# ======================================================================================
logging.debug('Create AutoML...')
automl = AutoML(reader=reader,
levels=[[gbm_lvl0, reg_lvl0]],
timer=timer,
blender=blender,
skip_conn=False)
logging.debug('AutoML created...')
# ======================================================================================
logging.debug('Fit predict...')
oof_pred = automl.fit_predict(train, roles={'target': "TARGET"})
logging.debug('Finished fitting...')
test_pred = automl.predict(test)
logging.debug('Prediction for test data:\n{}\nShape = {}'.format(
test_pred, test_pred.shape))
# ======================================================================================
logging.debug('Check scores...')
# use only not nan
not_nan = np.any(~np.isnan(oof_pred.data), axis=1)
logging.debug('OOF score: {}'.format(
log_loss(train['TARGET'].values[not_nan], oof_pred.data[not_nan])))
logging.debug('TEST score: {}'.format(
log_loss(test['TARGET'].values, test_pred.data)))
# ======================================================================================
for dat, df, name in zip([oof_pred, test_pred], [train, test],
['train', 'test']):
logging.debug('Check aucs {0}...'.format(name))
for c in range(3):
_sc = roc_auc_score((df['TARGET'].values == c).astype(np.float32),
dat.data[:, c])
logging.debug('Cl {0} auc score: {1}'.format(c, _sc))
def test_different_losses_and_metrics():
np.random.seed(42)
logging.basicConfig(format='[%(asctime)s] (%(levelname)s): %(message)s',
level=logging.DEBUG)
logging.debug('Load data...')
data = pd.read_csv('../example_data/test_data_files/sampled_app_train.csv')
logging.debug('Data loaded')
logging.debug('Features modification from user side...')
data['BIRTH_DATE'] = (
np.datetime64('2018-01-01') +
data['DAYS_BIRTH'].astype(np.dtype('timedelta64[D]'))).astype(str)
data['EMP_DATE'] = (np.datetime64('2018-01-01') +
np.clip(data['DAYS_EMPLOYED'], None, 0).astype(
np.dtype('timedelta64[D]'))).astype(str)
data['constant'] = 1
data['allnan'] = np.nan
data.drop(['DAYS_BIRTH', 'DAYS_EMPLOYED'], axis=1, inplace=True)
logging.debug('Features modification finished')
logging.debug('Split data...')
train_data, test_data = train_test_split(data,
test_size=2000,
stratify=data['TARGET'],
random_state=13)
train_data.reset_index(drop=True, inplace=True)
test_data.reset_index(drop=True, inplace=True)
logging.debug(
'Data splitted. Parts sizes: train_data = {}, test_data = {}'.format(
train_data.shape, test_data.shape))
for task_params, target in zip([{
'name': 'binary'
}, {
'name': 'binary',
'metric': roc_auc_score
}, {
'name': 'reg',
'loss': 'mse',
'metric': 'r2'
}, {
'name': 'reg',
'loss': 'rmsle',
'metric': 'rmsle'
}, {
'name': 'reg',
'loss': 'quantile',
'loss_params': {
'q': .9
},
'metric': 'quantile',
'metric_params': {
'q': .9
}
}], ['TARGET', 'TARGET', 'AMT_CREDIT', 'AMT_CREDIT', 'AMT_CREDIT']):
logging.debug('Create task..')
task = Task(**task_params)
logging.debug('Task created')
logging.debug('Create reader...')
reader = PandasToPandasReader(task, cv=5, random_state=1)
logging.debug('Reader created')
# pipeline 1 level parts
logging.debug('Start creation pipeline_1...')
pipe = LGBSimpleFeatures()
logging.debug('\t ParamsTuner2 and Model2...')
model2 = BoostLGBM(default_params={
'learning_rate': 0.025,
'num_leaves': 64,
'seed': 2,
'num_threads': 5
})
logging.debug('\t Tuner2 and model2 created')
logging.debug('\t Pipeline1...')
pipeline_lvl1 = MLPipeline(
[model2],
pre_selection=None, # selector,
features_pipeline=pipe,
post_selection=None)
logging.debug('Pipeline1 created')
logging.debug('Create AutoML pipeline...')
automl = AutoML(reader, [
[pipeline_lvl1],
], skip_conn=False)
logging.debug('AutoML pipeline created...')
logging.debug('Start AutoML pipeline fit_predict...')
start_time = time.time()
oof_pred = automl.fit_predict(train_data, roles={'target': target})
logging.debug(
'AutoML pipeline fitted and predicted. Time = {:.3f} sec'.format(
time.time() - start_time))
test_pred = automl.predict(test_data)
logging.debug('Prediction for test data:\n{}\nShape = {}'.format(
test_pred, test_pred.shape))
logging.debug('Check scores...')
logging.debug('OOF score: {}'.format(
task.metric_func(train_data[target].values, oof_pred.data[:, 0])))
logging.debug('TEST score: {}'.format(
task.metric_func(test_data[target].values, test_pred.data[:, 0])))
logging.debug('Pickle automl')
with open('automl.pickle', 'wb') as f:
pickle.dump(automl, f)
logging.debug('Load pickled automl')
with open('automl.pickle', 'rb') as f:
automl = pickle.load(f)
logging.debug('Predict loaded automl')
test_pred = automl.predict(test_data)
logging.debug('TEST score, loaded: {}'.format(
roc_auc_score(test_data['TARGET'].values, test_pred.data[:, 0])))
os.remove('automl.pickle')