def tune_gbdt_params(model, X, y, n_splits) -> dict:
'''
Tune parameters by training
'''
# start tuning train log
create_logger('train', **c.log)
logger_train = getLogger('train')
logger_train.debug('{}\t{}\t{}\t{}'.format('fold', 'iteration',
'train_auc', 'val_auc'))
aucs = list()
feature_importances = pd.DataFrame()
feature_importances['feature'] = X.columns
# split data into train, validation
folds = TimeSeriesSplit(n_splits=n_splits)
for i, (idx_train, idx_val) in enumerate(folds.split(X, y)):
fold = i + 1
with blocktimer(f'Fold {fold}'):
# prepare
logger.info(f'Training on fold {fold}')
X_train = X.iloc[idx_train]
y_train = y.iloc[idx_train]
X_val = X.iloc[idx_val]
y_val = y.iloc[idx_val]
# train
model = model.train_and_validate(X_train, y_train, X_val, y_val,
logger_train, fold)
# record result
feature_importances[f'fold_{fold}'] = model.feature_importance
aucs.append(model.validation_auc)
# TODO: save models at each steps
logger.debug(f'Fold {fold} finished')
logger.info('Training has finished.')
logger.debug(f'Mean AUC: {np.mean(aucs)}')
# TODO: save feature importance and other
# make optimal config from result
optimal_c_model = model.config
if model.best_iteration is not None:
# new param
optimal_c_model.params['num_boost_round'] = model.best_iteration
else:
logger.warn(
'Did not meet early stopping. Try larger num_boost_rounds.')
# no need after optimized num_boost_round
del optimal_c_model.params['early_stopping_rounds']
return optimal_c_model
def tune_gbdt_params(model, X, y, n_splits) -> dict:
'''
Tune parameter num_boost_round
'''
# start tuning train log
create_logger('train', **c.log)
logger_train = getLogger('train')
logger_train.debug('{}\t{}\t{}\t{}'.format('fold', 'iteration',
'train_auc', 'val_auc'))
# aucs = list()
# split data into train, validation
folds = TimeSeriesSplit(n_splits=n_splits)
for i, (idx_train, idx_val) in enumerate(folds.split(X, y)):
fold = i + 1
with blocktimer(f'Training on Fold {fold}'):
X_train = X.iloc[idx_train]
y_train = y.iloc[idx_train]
X_val = X.iloc[idx_val]
y_val = y.iloc[idx_val]
# train
model = model.train_and_validate(X_train, y_train, X_val, y_val,
logger_train, fold)
model.save(
c.model.dir /
f'model_{c.runtime.VERSION}_{c.model.TYPE}_fold{fold}.pkl')
# record result
# aucs.append(model.val_auc)
# logger.info(f'train_auc: {model.train_auc} val_auc: {model.val_auc}')
# logger.info(f'Mean AUC: {np.mean(aucs)}')
# make optimal config from result
optimal_c_model = model.config
if model.best_iteration is not None:
optimal_c_model.params['num_boost_round'] = model.best_iteration
else:
logger.warn(
'Did not meet early stopping. Try larger num_boost_rounds.')
# no need after optimized num_boost_round
del optimal_c_model.params['early_stopping_rounds']
return optimal_c_model
return optimal_c_model
if __name__ == "__main__":
gc.enable()
warnings.filterwarnings('ignore')
# read config & apply option
c = EasyDict(config)
opt = parse_option()
c.transformer.USE_SMALL_DATA = opt.small
c.log.slackauth.NO_SEND_MESSAGE = opt.nomsg
seed_everything(c.runtime.RANDOM_SEED)
create_logger('main', **c.log)
logger = getLogger('main')
logger.info(
f':thinking_face: Starting experiment {c.runtime.VERSION}_{c.runtime.DESCRIPTION}'
)
try:
main(c)
logger.info(
f':sunglasses: Finished experiment {c.runtime.VERSION}_{c.runtime.DESCRIPTION}'
)
except Exception:
logger.critical(
f':smiling_imp: Exception occured \n {traceback.format_exc()}')
logger.critical(
f':skull: Stopped experiment {c.runtime.VERSION}_{c.runtime.DESCRIPTION}'
seed_everything(c.runtime.random_seed)
# check config.experiment_type
if c.experiment_type != experiment_type:
raise ValueError(
f'experiment_type in config: {c.experiment_type} does not match this script'
)
# start logging
r.paths.main_log_path = f'log/main_{c.runtime.version}{c.runtime.dsize}.log'
r.paths.train_log_path = f'log/train_{c.runtime.version}{c.runtime.dsize}.tsv'
slackauth = EasyDict(json.load(open('./slackauth.json', 'r')))
slackauth.token_path = Path().home() / slackauth.token_file
create_logger('main',
version=c.runtime.version,
log_path=r.paths.main_log_path,
slackauth=slackauth,
no_send_message=c.runtime.no_send_message)
create_logger('train',
version=c.runtime.version,
log_path=r.paths.train_log_path,
slackauth=slackauth,
no_send_message=c.runtime.no_send_message)
logger = getLogger('main')
logger.info(
f':thinking_face: Starting experiment {c.runtime.version}_{c.model.type}{c.runtime.dsize}'
)
logger.info(f'Options indicated: {opt}')
'''
logger_train = getLogger('train')
logger_train.debug('{}\t{}\t{}\t{}'.format('fold', 'iteration', 'train_auc', 'val_auc'))