-
Notifications
You must be signed in to change notification settings - Fork 0
/
xgb_train.py
executable file
·158 lines (127 loc) · 5.52 KB
/
xgb_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
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
import xgboost as xgb
import numpy as np
import pandas as pd
import os
from sklearn.model_selection import KFold
from imblearn.over_sampling import SMOTE
dense_train_file = '/home/zydq/Datasets/LCZ/dense_f_train.csv'
dense_val_file = '/home/zydq/Datasets/LCZ/dense_f_val.csv'
dense_test_file = '/home/zydq/Datasets/LCZ/dense_f_test.csv'
NUM_ROUNDS = 100000
SEED = 502
FOLD = 3
GABOR=False
extra = ''
# resample strategy
VAL_RESAMPLE_TIMES = 10
if GABOR:
dense_train_file = '/home/zydq/Datasets/LCZ/dense_f_gabor_train.csv'
dense_val_file = '/home/zydq/Datasets/LCZ/dense_f_gabor_val.csv'
dense_test_file = '/home/zydq/Datasets/LCZ/dense_f_gabor_test.csv'
extra = 'gabor'
base_dir = './xgb_ckp_' + extra +str(FOLD)+ 'fold/'
if not os.path.isdir(base_dir):
os.mkdir(base_dir)
def gen_resample_dict(Y, times):
label, count = np.unique(Y, return_counts=True)
return dict(zip(label, (times * count).astype(int)))
if __name__ == '__main__':
train_df = pd.read_csv(dense_train_file)
val_df = pd.read_csv(dense_val_file)
test_df = pd.read_csv(dense_test_file)
feat_names = train_df.columns.tolist()[:-1]
train_X = train_df.values[:, :-1]
train_Y = train_df.values[:, -1]
val_X = val_df.values[:, :-1]
val_Y = val_df.values[:, -1]
val_idx = list(range(val_Y.shape[0]))
test_X = test_df.values
del train_df, val_df, test_df
# SMOTE
sm = SMOTE(random_state=SEED, sampling_strategy=lambda y: gen_resample_dict(y, VAL_RESAMPLE_TIMES))
d_test = xgb.DMatrix(test_X, feature_names=feat_names)
test_pred_total = None
val_pred_total = None
y_total = None
# trainval_Y = np.eye(17)[trainval_df['label'].astype(int).values.reshape(-1)]
kf = KFold(n_splits=FOLD, random_state=SEED, shuffle=True)
# fold_xgb = []
folds_stack_x = np.array([])
folds_stack_y = np.array([])
for foldid, (train_index, val_index) in enumerate(kf.split(train_X)):
print('=' * 80)
print('starting fold: ', foldid)
print('=' * 80)
np.random.seed(SEED)
np.random.shuffle(val_idx)
val_X = val_X[val_idx]
val_Y = val_Y[val_idx]
val_res_X, val_res_Y = sm.fit_resample(val_X, val_Y)
# train on big one
x_train, x_val = train_X[train_index], train_X[val_index]
y_train, y_val = train_Y[train_index], train_Y[val_index]
# train on small one
# x_val, x_train = train_X[train_index], train_X[val_index]
# y_val, y_train = train_Y[train_index], train_Y[val_index]
x_train = np.concatenate([x_train, val_res_X], axis=0)
y_train = np.concatenate([y_train, val_res_Y], axis=0)
# shuffle
train_shuffled_idx = list(range(x_train.shape[0]))
np.random.seed(SEED)
np.random.shuffle(train_shuffled_idx)
x_train, y_train = x_train[train_shuffled_idx], y_train[train_shuffled_idx]
print((x_train.shape))
params = {
'tree_method': 'gpu_hist',
'booster': 'gbtree',
'objective': 'multi:softprob', # 多分类的问题
'eval_metric': 'merror',
'num_class': 17, # 类别数，与 multisoftmax 并用
'gamma': 2, # 用于控制是否后剪枝的参数,越大越保守，一般0.1、0.2这样子。
'max_depth': 4, # 构建树的深度，越大越容易过拟合
'n_estimators': 15,
'lambda': 8, # L2
'alpha': 8, # L1
'subsample': 0.5, # 随机采样训练样本
'colsample_by*': 0.5, # 生成树时进行的列采样
'min_child_weight': 5,
'scale_pos_weight': 5,
'silent': 1, # 设置成1则没有运行信息输出，最好是设置为0.
'eta': 0.02, # 如同学习率
'seed': 502,
# 'nthread': 4, # cpu 线程数
}
d_train = xgb.DMatrix(x_train, label=y_train, feature_names=feat_names)
d_val = xgb.DMatrix(x_val, label=y_val, feature_names=feat_names)
watchlist = [(d_train, 'train'), (d_val, 'valid')]
bst_xgb = xgb.train(params, d_train, NUM_ROUNDS, watchlist, early_stopping_rounds=10, verbose_eval=10)
# fold_xgb.append(bst_xgb)
val_pred = bst_xgb.predict(d_val).reshape(y_val.shape[0], 17)
val_score = (np.argmax(val_pred, -1) == y_val).sum() / y_val.shape[0]
print("val score :", val_score)
if val_pred_total is None:
val_pred_total = val_pred
y_total = y_val
else:
val_pred_total = np.concatenate([val_pred_total, val_pred], axis=0)
y_total = np.concatenate([y_total, y_val], axis=0)
test_pred = bst_xgb.predict(d_test).reshape(test_X.shape[0], 17, 1)
if test_pred_total is None:
test_pred_total = test_pred
else:
test_pred_total = np.concatenate([test_pred_total, test_pred], axis=-1)
fold_submit = np.eye(17)[test_pred.argmax(1).reshape(-1)]
np.savetxt('./submit/sub_xgb_'+ extra +str(FOLD)+ 'fold' + str(foldid) + '_' + str(val_score) + '.csv', fold_submit, delimiter=',', fmt='%d')
np.savetxt('./score/score_xgb_'+ extra +str(FOLD)+ 'fold' + str(foldid) + '_' + str(val_score) + '.csv', test_pred[:, :, 0], delimiter=',', fmt='%.5f')
bst_xgb.save_model(os.path.join(base_dir, 'xgb_'+ extra +str(FOLD)+ 'fold' + str(foldid) + '.model'))
bst_xgb.dump_model(os.path.join(base_dir, 'xgb_'+ extra +str(FOLD)+ 'fold' + str(foldid) + '.dump.raw'))
bst_xgb.__del__()
d_train.__del__()
d_val.__del__()
del d_train, d_val, x_train, x_val, y_train, y_val, val_res_X, val_res_Y
total_score = (np.argmax(val_pred_total, -1) == y_total).sum() / y_total.shape[0]
print("total val score :", total_score)
test_pred_total = test_pred_total.mean(axis=-1, keepdims=False)
submit = np.eye(17)[test_pred_total.argmax(-1).reshape(-1)]
np.savetxt('./submit/sub_xgb_' + extra + str(FOLD)+ 'full_' + str(total_score) + '.csv', submit, delimiter=',', fmt='%d')
np.savetxt('./score/score_xgb_' + extra + str(FOLD)+ 'full_' + str(total_score) + '.csv', test_pred_total, delimiter=',', fmt='%.5f')