在此之前，调参要么网格调参，要么随机调参，要么肉眼调参。虽然调参到一定程度，进步有限，但仍然很耗精力。
自动调参库hyperopt可用tpe算法自动调参，实测强于随机调参。
hyperopt 需要自己写个输入参数，返回模型分数的函数（只能求最小化，如果分数是求最大化的，加个负号），设置参数空间。
本来最优参数fmin函数会自己输出的，但是出了意外，参数会强制转化整数，没办法只好自己动手了。
demo如下：
import lightgbm as lgb
from sklearn.metrics import roc_auc_score as auc
def get_set(n1,data='trained.csv',n_splits=10,y=False,random_state=0):
from sklearn.model_selection import KFold
data=pd.read_csv(data)
kf = KFold(n_splits=n_splits,shuffle=True,random_state=random_state)
if y:
train,test=pd.DataFrame(),pd.DataFrame()
clas=list(data[y].unique())
for cla in clas:
i=0
dd=data[data[y]==cla]
for train_index,test_index in kf.split(dd):
i=i+1
if n1==i:
train=train.append(data.loc[list(train_index)])
test=test.append(data.loc[list(test_index)])
else:
i=0
for train_index,test_index in kf.split(data):
i=i+1
if n1==i:
train=data.iloc[list(train_index),:]
test=data.iloc[list(test_index),:]
return train,test
def scorer(yp,data):
yt= data.get_label()
score=auc(yt,yp)
return 'auc',score,True
def peropt(param):
conf=['num_leaves','max_depth','min_child_samples','max_bin']
for i in conf:
param[i]=int(param[i])
evals_result={}
lgb.train(param,
dtrain,
2000,
feval=scorer,
valid_sets=[dval],
verbose_eval=None,
evals_result=evals_result,
early_stopping_rounds=10)
best_score=evals_result['valid_0']['auc'][-11]
#print(param,best_score,len(evals_result['valid_0']['auc'])-10)
result.append((param,best_score,len(evals_result['valid_0']['auc'])-10))
return -best_score
if 0:#数据集
i=1
x_train,x_test=get_set(i,n_splits=5)
x_train.pop('CaseId')
x_test.pop('CaseId')
y_train=x_train.pop('Evaluation')
y_test=x_test.pop('Evaluation')
dtrain=lgb.Dataset(x_train,y_train)
dval=lgb.Dataset(x_test,y_test)
if 1:#调参
from hyperopt import fmin,tpe,hp#,rand#,pyll#,partial
space={ 'num_leaves': hp.quniform('num_leaves',50,70,1)
,'max_depth':hp.quniform('max_depth',7,15,1)
,'min_child_samples':hp.quniform('min_child_samples',5,20,1)
,'max_bin':hp.quniform('max_bin',100,150,5)
,'learning_rate':hp.choice('learning_rate',[0.01])
,'subsample':hp.uniform('subsample',0.9,1)
,'colsample_bytree':hp.uniform('colsample_bytree',0.95,1)
,'min_split_gain':hp.loguniform('min_split_gain',-5,2)
,'reg_alpha':hp.loguniform('reg_alpha',-5,2)
,'reg_lambda':hp.loguniform('reg_lambda',-5,2)
}
result=[]
#print(pyll.stochastic.sample(space))#抽样
#algo=partial(tpe.suggest,n_startup_jobs=10)#作用未知
fmin(peropt,
space=space,
algo=tpe.suggest,
max_evals=100
)
sort=sorted(result,key=lambda x:x[1],reverse=True)