This repository has been archived by the owner on Apr 17, 2021. It is now read-only.
/
run_lightgbm.py
237 lines (176 loc) · 7.63 KB
/
run_lightgbm.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
import gc
import time
import warnings
import yaml
from argparse import ArgumentParser
from pathlib import Path
import joblib
import logzero
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from joblib import Parallel, delayed
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import StratifiedKFold
from _model_map import models
from _feature_map import features
from utils import fast_auc, load_train_dataset, load_test_dataset, load_dump, timer
def get_dataset_filename(config, dataset_type):
assert (type(dataset_type) == str and
dataset_type in ('train', 'target', 'test', 'sample_submit'))
dataset_dir = Path(config['dataset']['input_directory'])
filename = Path(config['dataset']['files'][dataset_type])
return dataset_dir / filename
def setup_logger(output_dir, logname):
return logzero.setup_logger(
logfile=f'{output_dir}/{logname}.log',
level=10,
formatter=None,
)
def create_features(config, train_df, test_df, kfold):
def _process(feature):
feat = get_feature(feature)
encode_type = None
if feature in config['features']['count_encode']:
encode_type = 'count'
elif feature in config['features']['category_encode']:
encode_type = 'category'
elif feature in config['features']['rank_encode']:
encode_type = 'rank'
elif feature in config['features']['target_encode']:
encode_type = 'target'
feat.prepare_target(config['target'])
if feat.create(train_df, test_df, kfold, encode_type):
feat.save()
with timer('Create features...'):
Parallel(n_jobs=1)([delayed(_process)(f) for f in config['features']['all']])
def get_feature(feature):
return features[feature]()
def get_kfold(config):
cv = config['cv_strategy']
if cv['method'] == 'StratifiedKFold':
kfold = StratifiedKFold(n_splits=cv['n_splits'],
shuffle=True,
random_state=cv['seed'])
else:
raise NotImplementedError
return kfold
def extract_use_features(config):
use_features = []
for f in config['features']['all']:
if f in config['features']['count_encode']:
use_features.append(f'cnt_{f}')
elif f in config['features']['category_encode']:
use_features.append(f'lbl_{f}')
elif f in config['features']['rank_encode']:
use_features.append(f'rank_{f}')
elif f in config['features']['target_encode']:
cv_n_splits = config['cv_strategy']['n_splits']
cv_seed = config['cv_strategy']['seed']
use_features.append(f'te_{f}_split{cv_n_splits}_seed{cv_seed}')
else:
use_features.append(f)
return use_features
def train_model(x_train, y_train, kfold, config, logger):
feature_importances = pd.DataFrame()
scores = []
clfs = []
for i, (train_index, valid_index) in enumerate(kfold.split(x_train, y_train)):
start_time = time.time()
logger.info(f'Fold {i+1}')
x_train_fold = x_train.iloc[train_index]
y_train_fold = y_train.iloc[train_index]
x_valid_fold = x_train.iloc[valid_index]
y_valid_fold = y_train.iloc[valid_index]
clf = models[config['model']['name']]()
y_pred_valid = clf.train_and_validate(x_train_fold,
x_valid_fold,
y_train_fold,
y_valid_fold,
params=config['model'],
logger=logger)
scores.append(fast_auc(y_valid_fold, y_pred_valid))
logger.info(f'Fold roc_auc: {roc_auc_score(y_valid_fold.values, y_pred_valid)}')
fold_importance = pd.DataFrame()
fold_importance['feature'] = x_train_fold.columns
fold_importance['gain'] = clf.model.feature_importance(importance_type='gain')
fold_importance['fold'] = i + 1
feature_importances = pd.concat([feature_importances, fold_importance],
axis=0, sort=False)
clfs.append(clf)
# resume training when error occurs
joblib.dump(clf, f'tmp/lgb_fold{i+1}.pkl')
joblib.dump(feature_importances, f'tmp/importances_fold{i+1}.pkl')
elapsed = time.time() - start_time
logger.info(f'Fold {i+1} done in {elapsed:.0f} s\n')
feature_importances['gain'] /= kfold.n_splits
logger.info(f'CV mean score: {np.mean(scores):.4f}')
return clfs, feature_importances
def predict_model(clfs, use_features, config):
n_splits = len(clfs)
predictions = []
for i, clf in enumerate(clfs):
x_test = load_test_dataset(use_features)
for f in config['features']['target_encode']:
if f is None:
pass
col_name = get_feature(f).name
drop_cols = [f'te_{col_name}_fold{k+1}' for k in range(n_splits) if i != k]
x_test.drop(drop_cols, axis=1, inplace=True)
x_test.rename(columns={f'te_{col_name}_fold{i+1}': f'te_{col_name}'},
inplace=True)
print(f'predict: fold {i+1}/{n_splits}')
y_pred = clf.predict(x_test)
predictions.append(y_pred)
return np.mean(predictions, axis=0)
def save_feature_importances(importances, basename, output_dir):
mean_gain = importances[['feature', 'gain']].groupby('feature').mean()
importances['mean_gain'] = importances['feature'].map(mean_gain['gain'])
with warnings.catch_warnings():
warnings.simplefilter('ignore')
plt.figure(figsize=(20, 30))
sns.barplot(x='gain', y='feature',
data=importances.sort_values('mean_gain', ascending=False))
plt.title('LightGBM feature importances')
plt.tight_layout()
plt.savefig(f'{output_dir}/importances_{basename}.png')
def main():
parser = ArgumentParser()
parser.add_argument('--config', default='./configs/lgb_template.yaml')
parser.add_argument('--create-features', action='store_true')
options = parser.parse_args()
config = yaml.safe_load(open(options.config))
kfold = get_kfold(config)
if options.create_features:
train_path = get_dataset_filename(config, 'train')
test_path = get_dataset_filename(config, 'test')
with timer('Load train/test dump files'):
train_df = load_dump(train_path)
test_df = load_dump(test_path)
create_features(config, train_df, test_df, kfold)
del train_df, test_df
gc.collect()
target_col = config['target']
target_path = get_dataset_filename(config, 'target')
use_features = extract_use_features(config)
x_train = load_train_dataset(use_features)
y_train = load_dump(target_path)[target_col]
output_dir = config['dataset']['output_directory']
basename = Path(options.config).stem
logger = setup_logger(output_dir, basename)
clfs, importances = train_model(x_train, y_train, kfold, config, logger)
save_feature_importances(importances, basename, output_dir)
del x_train, y_train, importances
gc.collect()
pred = predict_model(clfs, use_features, config)
print('Creating a submission csv file...')
submission_path = get_dataset_filename(config, 'sample_submit')
submission = pd.read_csv(submission_path)
submission[target_col] = pred
submission.to_csv(f'{output_dir}/submit_{basename}.csv.gz', index=False)
print('Done.')
if __name__ == '__main__':
main()