基于Python贝叶斯优化XGBoost算法调参报错“TypeError: ‘float‘ object is not subscriptable”

Posted 2022-12-04 肖永威

基于Python贝叶斯优化XGBoost算法调参运行情况如下：

iter	target	colsam…	gamma	max_depth	min_ch…	subsample
1	0.9398	0.8043	0.7483	6.0	276.7	0.6514
2	0.9405	0.7231	0.2676	7.382	347.6	0.7886
3	0.9388	0.8048	0.7167	6.818	708.6	0.6096
4	0.9421	0.8676	0.4756	8.235	155.3	0.6693
5	0.9399	0.9002	0.9714	7.254	569.2	0.9067

报出如下错误：

Traceback (most recent call last):
......
    suggestion = acq_max(
  File "/usr/local/python3/lib/python3.8/site-packages/bayes_opt/util.py", line 65, in acq_max
    if max_acq is None or -res.fun[0] >= max_acq:
TypeError: 'float' object is not subscriptable

参考关键代码如下：

def _xgb_logistic_evaluate(max_depth, subsample, gamma, colsample_bytree, min_child_weight):
    import xgboost as xgb

    params = 
        'objective': 'binary:logistic',  # 逻辑回归二分类的问题
        'eval_metric': 'auc',
        'max_depth': int(max_depth),
        'subsample': subsample,  # 0.8
        'eta': 0.3,
        'gamma': gamma,
        'colsample_bytree': colsample_bytree,
        'min_child_weight': min_child_weight

    cv_result = xgb.cv(params, self.dtrain,
                       num_boost_round=30, nfold=5)

    return 1.0 * cv_result['test-auc-mean'].iloc[-1]
    
def evaluate(self, bo_f, pbounds, init_points, n_iter):

    bo = BayesianOptimization(
        f=bo_f,   # 目标函数
        pbounds=pbounds,  # 取值空间
        verbose=2,  # verbose = 2 时打印全部，verbose = 1 时打印运行中发现的最大值，verbose = 0 将什么都不打印
        random_state=1,
        )

    bo.maximize(init_points=init_points,   # 随机搜索的步数
                n_iter=n_iter,       # 执行贝叶斯优化迭代次数
                acq='ei')

    print(bo.max)
    res = bo.max
    params_max = res['params']

    return params_max

参考stackoverflow上的解释：

This is related to a change in scipy 1.8.0, One should use -np.squeeze(res.fun) instead of -res.fun[0]

https://github.com/fmfn/BayesianOptimization/issues/300

The comments in the bug report indicate reverting to scipy 1.7.0 fixes this,

UPDATED: It seems the fix has been merged in the BayesianOptimization package, but the new maintainer is unable to push a release to pypi https://github.com/fmfn/BayesianOptimization/issues/300#issuecomment-1146903850

因此，卸载当前scipy 1.8.1，退回到scipy 1.7.0。

[root@DeepLearning bin]# pip3 uninstall scipy
......
  Successfully uninstalled scipy-1.8.1
[root@DeepLearning bin]# pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple scipy==1.7
  Successfully installed scipy-1.7.0

成功再运行贝叶斯优化调参程序。

参考：

seul233. python使用贝叶斯优化随机森林时出现TypeError: ‘float’ object is not subscriptable. CSDN博客. 2022.03

https://stackoverflow.com/questions/71460894/bayesianoptimization-fails-due-to-float-error