-
Notifications
You must be signed in to change notification settings - Fork 21
/
cbt_train.py
132 lines (105 loc) · 4.77 KB
/
cbt_train.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
# -*- coding：utf-8 -*-
# Author: juzstu
# Time: 2021/8/7 18:19
from catboost import CatBoostClassifier, Pool
import pandas as pd
from sklearn.metrics import roc_auc_score
import numpy as np
import os
import logging
logger = logging.getLogger()
formatter = logging.Formatter('%(asctime)s - %(message)s')
fhandler = logging.FileHandler('cbt_train.log', 'w')
fhandler.setLevel(logging.DEBUG)
fhandler.setFormatter(formatter)
chandler = logging.StreamHandler()
chandler.setLevel(logging.INFO)
chandler.setFormatter(formatter)
logger.addHandler(fhandler)
logger.addHandler(chandler)
logger.setLevel(logging.INFO)
def calc_uauc(df, label):
df = df.groupby('userid')[[label, f'pred_{label}']].agg(list).reset_index()
df['auc'] = df.apply(lambda x: roc_auc_score(x[label], x[f'pred_{label}']), axis=1)
return df['auc'].mean()
def single_cbt_train(train_: pd.DataFrame, valid_: pd.DataFrame, use_train_feats: list, label: str, eta: float):
print('data shape:\ntrain--{}\nvalid--{}\n'.format(train_.shape, valid_.shape))
print('Use {} features ...'.format(len(use_train_feats)))
print('Use catboost to train ...')
params = {
'iterations': 50000,
'learning_rate': eta,
'max_depth': 6,
'random_seed': 1,
'eval_metric': 'AUC',
'task_type': 'GPU',
'early_stopping_rounds': 1000,
'use_best_model': True,
'verbose': 100
}
train_pool = Pool(train_[use_train_feats], train_[label], feature_names=use_train_feats)
valid_pool = Pool(valid_[use_train_feats], valid_[label], feature_names=use_train_feats)
clf = CatBoostClassifier(**params)
clf.fit(train_pool, eval_set=valid_pool)
imp_df = pd.DataFrame()
imp_df['feats'] = clf.feature_names_
imp_df['imp'] = clf.feature_importances_
imp_df.sort_values('imp', ascending=False, inplace=True)
print(imp_df.head(50))
valid_['pred_{}'.format(label)] = clf.predict_proba(valid_[use_train_feats])[:, 1]
tmp_auc = calc_uauc(valid_, label)
print(f'UAUC of {label}: {tmp_auc}')
return tmp_auc, clf
def online_cbt_train(train_: pd.DataFrame, test_: pd.DataFrame, use_train_feats: list, label: str, eta: float,
n_rounds: int):
print('data shape:\ntrain--{}\ntest--{}\n'.format(train_.shape, test_.shape))
print('Use {} features ...'.format(len(use_train_feats)))
print('Use catboost to train ...')
params = {
'iterations': n_rounds,
'learning_rate': eta,
'max_depth': 6,
'random_seed': 1,
'task_type': 'GPU',
'verbose': 100
}
train_pool = Pool(train_[use_train_feats], train_[label], feature_names=use_train_feats)
clf = CatBoostClassifier(**params)
clf.fit(train_pool)
test_[label] = clf.predict_proba(test_[use_train_feats])[:, 1]
test_[['userid', 'feedid', label]].to_csv(f'{sub_path}/r2_testa_cbt_{label}.csv', index=False, encoding='utf8')
clf.save_model(f'{model_path}/cbt_{label}.cbm')
if __name__ == "__main__":
sub_path = './sub'
feat_path = './feats'
model_path = './tree_model'
label_cols = ['like', 'click_avatar', 'forward', 'follow', 'favorite', 'read_comment', 'comment']
rounds_dict = {}
auc_score = []
for n in label_cols:
logger.info(f'start train {n} ...')
train_set = pd.read_feather(f'{feat_path}/for_train_{n}.feather')
valid_set = pd.read_feather(f'{feat_path}/for_valid_{n}.feather')
use_cols = [i for i in valid_set.columns if i not in ['feedid', 'userid', 'date_'] + label_cols]
auc_, cbt_model = single_cbt_train(train_set, valid_set, use_cols, n, 0.05)
auc_score.append(auc_)
rounds_dict[n] = cbt_model.get_best_iteration()
logger.info('#' * 100)
for n in zip(label_cols, auc_score):
logger.info(f'AUC of {n[0]}: {n[1]}, rounds-{rounds_dict[n[0]]}')
logger.info(f'score: {np.average(auc_score, weights=[3, 2, 1, 1, 1, 4, 1])}')
test_set = pd.read_feather(feat_path + '/round2_test_feats.feather')
for n in label_cols:
now_model_path = f'{model_path}/cbt_{n}.cbm'
train_set = pd.read_feather(f'{feat_path}/all_train_{n}.feather')
use_cols = [i for i in train_set.columns if i not in ['feedid', 'userid', 'date_'] + label_cols]
if os.path.exists(now_model_path):
logger.info(f'find exist cbt model of {n}.')
model = CatBoostClassifier().load_model(now_model_path)
del train_set
test_set[n] = model.predict_proba(test_set[use_cols])[:, 1]
else:
logger.info(f'start train {n} ...')
online_cbt_train(train_set, test_set, use_cols, n, 0.05, rounds_dict[n]+1000)
logger.info('#' * 100)
test_set[['userid', 'feedid'] + label_cols].to_csv(f'{sub_path}/r2_testa_cbt_baseline.csv', index=False, encoding='utf8')