刷分神器，使用hyperopt实现lightgbm自动化调参！

lyhue1991

发布于 2021-08-06 15:30:19

2.1K0

发布于 2021-08-06 15:30:19

文章被收录于专栏：Python与算法之美Python与算法之美

Hyperopt是最受欢迎的调参工具包，目前在github上已经获得star数量5.8k，在kaggle天池等比赛中经常出现它的身影。

它的主要功能是应用随机搜索，模拟退火以及贝叶斯优化等优化算法，在不可解析不可求导的参数空间中求解函数的最小值。

下面我们首先来看看它的一些基本范例用法，最后再使用它来实现对lightgbm模型的自动化调参！

公众号算法美食屋后台回复关键词：hyperopt，获取本文全部代码对应的jupyter notebook 文件！?

一，单一参数空间

单一参数可以直接用 hyperopt.hp 中的相关参数类型指定搜索范围。

import numpy as np 
import matplotlib.pyplot as plt 
from hyperopt import fmin, tpe, hp, Trials,space_eval

# 1，定义目标函数
def loss(x):
    return (x-1)**2 
  
# 2，定义超参空间

#常用的搜索空间
#hp.choice 离散值
#hp.uniform 均匀分布
#hp.normal 正态分布
spaces = hp.uniform("x",-5.0,5.0)

# 3，执行搜索过程
# hyperopt支持如下搜索算法
#随机搜索(hyperopt.rand.suggest)
#模拟退火(hyperopt.anneal.suggest)
#TPE算法（hyperopt.tpe.suggest，贝叶斯优化，算法全称为Tree-structured Parzen Estimator Approach）
trials = Trials()
best = fmin(fn=loss, space=spaces, algo=tpe.suggest, max_evals=1000,trials=trials)
   
# 4，获取最优参数

best_params = space_eval(spaces,best)
print("best_params = ",best_params)


# 5，绘制搜索过程
losses = [x["result"]["loss"] for x in trials.trials]
minlosses = [np.min(losses[0:i+1]) for i in range(len(losses))] 
steps = range(len(losses))

fig,ax = plt.subplots(figsize=(6,3.7),dpi=144)
ax.scatter(x = steps, y = losses, alpha = 0.3)
ax.plot(steps,minlosses,color = "red",axes = ax)
plt.xlabel("step")
plt.ylabel("loss")
plt.yscale("log")

二，网格参数空间

多个参数可以用字典表述成网格参数空间形式。

import numpy as np 
import matplotlib.pyplot as plt 
from hyperopt import fmin, tpe, hp,anneal, Trials,space_eval

# 1，定义目标函数
def loss(params):
    x,y = params["x"],params["y"]
    return x**2+y**2

    
# 2，定义超参空间
hspaces = {"x":hp.uniform("x",-1.0,1.0),
          "y":hp.uniform("y",1.0,2.0)}


# 3，执行搜索过程
# hyperopt支持如下搜索算法
#随机搜索(hyperopt.rand.suggest)
#模拟退火(hyperopt.anneal.suggest)
#TPE算法（hyperopt.tpe.suggest，贝叶斯优化，算法全称为Tree-structured Parzen Estimator Approach）
trials = Trials()
best = fmin(fn=loss, space=hspaces, algo=anneal.suggest, max_evals=1000,trials=trials)
   
# 4，获取最优参数
best_params = space_eval(hspaces,best)
print("best_params = ",best_params)


# 5，绘制搜索过程
losses = [x["result"]["loss"] for x in trials.trials]
minlosses = [np.min(losses[0:i+1]) for i in range(len(losses))] 
steps = range(len(losses))

fig,ax = plt.subplots(figsize=(6,3.7),dpi=144)
ax.scatter(x = steps, y = losses, alpha = 0.3)
ax.plot(steps,minlosses,color = "red",axes = ax)
plt.xlabel("step")
plt.ylabel("loss")
plt.yscale("log")

三，树形参数空间

有时候，后面的参数依赖于之前一些参数的取值，可以用hyperopt.hp.choice表述成树形参数空间。


import math 
from hyperopt import fmin, tpe, hp, Trials,space_eval

# 1，定义目标函数
def loss(params):
    f = params[0]
    if f=="sin":
        x = params[1]["x"]
        return math.sin(x)**2
    elif f=="cos":
        x = params[1]["x"]
        y = params[1]["y"]
        return math.cos(x)**2+y**2
    elif f=="sinh":
        x = params[1]["x"]
        return math.sinh(x)**2
    else:
        assert f=="cosh"
        x = params[1]["x"]
        y = params[1]["y"]
        return math.cosh(x)**2+y**2
    
# 2，定义超参空间
spaces = hp.choice("f",
        [("sin",{"x":hp.uniform("x1",-math.pi/2,math.pi)}),
        ("cos",{"x":hp.uniform("x2",-math.pi/2,math.pi),"y":hp.uniform("y2",-1,1)}),
        ("sinh",{"x":hp.uniform("x3",-5,5)}),
        ("cosh",{"x":hp.uniform("x4",-5,5),"y":hp.uniform("y4",-1,1)})])

# 3，执行搜索过程

# hyperopt支持如下搜索算法
#随机搜索(hyperopt.rand.suggest)
#模拟退火(hyperopt.anneal.suggest)
#TPE算法（hyperopt.tpe.suggest，贝叶斯优化，算法全称为Tree-structured Parzen Estimator Approach）
trials = Trials()
best = fmin(fn=loss, space=spaces, algo=tpe.suggest, max_evals=1000,trials=trials)
   
# 4，获取最优参数
best_params = space_eval(spaces,best)
print("best_params = ",best_params)

# 5，绘制搜索过程
losses = [x["result"]["loss"] for x in trials.trials]
minlosses = [np.min(losses[0:i+1]) for i in range(len(losses))] 
steps = range(len(losses))

fig,ax = plt.subplots(figsize=(6,3.7),dpi=144)
ax.scatter(x = steps, y = losses, alpha = 0.3)
ax.plot(steps,minlosses,color = "red",axes = ax)
plt.xlabel("step")
plt.ylabel("loss")
plt.yscale("log")

四，LightGBM手动调参

下面我们将应用hyperopt来对lightgbm模型进行超参数调参。我们使用的是网格参数空间。

作为对比，我们先看看手动调9组参数的结果。

手动调参的范例代码如下。

我们分别尝试以下9组参数：

最优超参数组合如下

learning_rate = 0.05 # 分别尝试(0.1,0.05,0.01)
boosting_type = 'gbdt' # 分别尝试('gbdt','rf','dart')
num_leaves = 63 # 分别尝试(15,31,63)

最优验证集得分为 f1_score=0.96591，看起来还不错！?

train f1_score: 0.99448 
valid f1_score: 0.96591

import datetime
import numpy as np
import pandas as pd
import lightgbm as lgb
from sklearn import datasets
from sklearn.model_selection import train_test_split
from sklearn.metrics import f1_score
import warnings 
warnings.filterwarnings('ignore')

def printlog(info):
    nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
    print("\n"+"=========="*8 + "%s"%nowtime)
    print(info+'...\n\n')

#================================================================================
# 一，读取数据
#================================================================================
printlog("step1: reading data...")

# 读取dftrain,dftest
breast = datasets.load_breast_cancer()
df = pd.DataFrame(breast.data,columns = [x.replace(' ','_') for x in breast.feature_names])
df['label'] = breast.target
df['mean_radius'] = df['mean_radius'].apply(lambda x:int(x))
df['mean_texture'] = df['mean_texture'].apply(lambda x:int(x))
dftrain,dftest = train_test_split(df)

categorical_features = ['mean_radius','mean_texture']
lgb_train = lgb.Dataset(dftrain.drop(['label'],axis = 1),label=dftrain['label'],
                        categorical_feature = categorical_features)

lgb_valid = lgb.Dataset(dftest.drop(['label'],axis = 1),label=dftest['label'],
                        categorical_feature = categorical_features,
                        reference=lgb_train)

#================================================================================
# 二，设置参数
#================================================================================
printlog("step2: setting parameters...")
                               
boost_round = 100                   
early_stop_rounds = 50

params = {
    'learning_rate': 0.05,
    'boosting_type': 'gbdt', #'dart','rf' 
    'num_leaves':63,
    'objective':'binary',
    'metric': ['auc'],
    'feature_fraction': 0.9,
    'bagging_fraction': 0.8,
    'bagging_freq': 5,
    'verbose': -1
}

#================================================================================
# 三，训练模型
#================================================================================
printlog("step3: training model...")


results = {}
gbm = lgb.train(params,
                lgb_train,
                num_boost_round= boost_round,
                valid_sets=(lgb_valid, lgb_train),
                valid_names=('validate','train'),
                early_stopping_rounds = early_stop_rounds,
                evals_result= results,
                verbose_eval = True)

#================================================================================
# 四，评估模型
#================================================================================
printlog("step4: evaluating model ...")

y_pred_train = gbm.predict(dftrain.drop('label',axis = 1), num_iteration=gbm.best_iteration)
y_pred_test = gbm.predict(dftest.drop('label',axis = 1), num_iteration=gbm.best_iteration)

train_score = f1_score(dftrain['label'],y_pred_train>0.5)
val_score = f1_score(dftest['label'],y_pred_test>0.5)

print('train f1_score: {:.5} '.format(train_score))
print('valid f1_score: {:.5} \n'.format(val_score))

lgb.plot_metric(results)
lgb.plot_importance(gbm,importance_type = "gain")

#================================================================================
# 五，保存模型
#================================================================================
printlog("step5: saving model ...")


model_dir = "gbm.model"
print("model_dir: %s"%model_dir)
gbm.save_model("gbm.model",num_iteration=gbm.best_iteration)
printlog("task end...")

###
##
#

五，LightGBM自动化调参

下面我们利用hyperopt对lightgbm进行自动化调参，我们一共尝试一百个超参组合。

以下程序用时较长，可以根据情况增加或者减少尝试的超参数组合个数。

注意我们的num_boost_round是通过early_stop自适应的，无需调参。

import datetime
import numpy as np
import pandas as pd
import lightgbm as lgb
from sklearn import datasets
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score,f1_score
import matplotlib.pyplot as plt 
from hyperopt import fmin,hp,Trials,space_eval,rand,tpe,anneal
import warnings 
warnings.filterwarnings('ignore')

def printlog(info):
    nowtime = datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')
    print("\n"+"=========="*8 + "%s"%nowtime)
    print(info+'...\n\n')

#================================================================================
# 一，读取数据
#================================================================================
printlog("step1: reading data...")

# 读取dftrain,dftest
breast = datasets.load_breast_cancer()
df = pd.DataFrame(breast.data,columns = [x.replace(' ','_') for x in breast.feature_names])
df['label'] = breast.target
df['mean_radius'] = df['mean_radius'].apply(lambda x:int(x))
df['mean_texture'] = df['mean_texture'].apply(lambda x:int(x))
dftrain,dftest = train_test_split(df)

categorical_features = ['mean_radius','mean_texture']
lgb_train = lgb.Dataset(dftrain.drop(['label'],axis = 1),label=dftrain['label'],
                        categorical_feature = categorical_features,free_raw_data=False)

lgb_valid = lgb.Dataset(dftest.drop(['label'],axis = 1),label=dftest['label'],
                        categorical_feature = categorical_features,
                        reference=lgb_train,free_raw_data=False)

#================================================================================
# 二，搜索超参
#================================================================================
printlog("step2: searching parameters...")

boost_round = 1000                   
early_stop_rounds = 50

params = {
    'learning_rate': 0.1,
    'boosting_type': 'gbdt',#'dart','rf'  
    'objective':'binary',
    'metric': ['auc'],
    'num_leaves': 31,
    'max_depth':  6,
    'min_data_in_leaf': 5,  
    'min_gain_to_split': 0,
    'reg_alpha':0,
    'reg_lambda':0,
    'feature_fraction': 0.9,
    'bagging_fraction': 0.8,
    'bagging_freq': 5,
    'feature_pre_filter':False,
    'verbose': -1
}

# 1，定义目标函数
def loss(config):
    params.update(config)
    gbm = lgb.train(params,
                    lgb_train,
                    num_boost_round= boost_round,
                    valid_sets=(lgb_valid, lgb_train),
                    valid_names=('validate','train'),
                    early_stopping_rounds = early_stop_rounds,
                    verbose_eval = False)
    y_pred_test = gbm.predict(dftest.drop('label',axis = 1), num_iteration=gbm.best_iteration)
    val_score = f1_score(dftest['label'],y_pred_test>0.5)

    return -val_score

  
# 2，定义超参空间

#可以根据需要，注释掉偏后的一些不太重要的超参
spaces = {"learning_rate":hp.loguniform("learning_rate",np.log(0.001),np.log(0.5)),
          "boosting_type":hp.choice("boosting_type",['gbdt','dart','rf']),
          "num_leaves":hp.choice("num_leaves",range(15,128)),
          #"max_depth":hp.choice("max_depth",range(3,11)),
          #"min_data_in_leaf":hp.choice("min_data_in_leaf",range(1,50)),
          #"min_gain_to_split":hp.uniform("min_gain_to_split",0.0,1.0),
          #"reg_alpha": hp.uniform("reg_alpha", 0, 2),
          #"reg_lambda": hp.uniform("reg_lambda", 0, 2),
          #"feature_fraction":hp.uniform("feature_fraction",0.5,1.0),
          #"bagging_fraction":hp.uniform("bagging_fraction",0.5,1.0),
          #"bagging_freq":hp.choice("bagging_freq",range(1,20))
          }


# 3，执行超参搜索
# hyperopt支持如下搜索算法
#随机搜索(hyperopt.rand.suggest)
#模拟退火(hyperopt.anneal.suggest)
#TPE算法（hyperopt.tpe.suggest，算法全称为Tree-structured Parzen Estimator Approach）
trials = Trials()
best = fmin(fn=loss, space=spaces, algo= tpe.suggest, max_evals=100, trials=trials)
   
# 4，获取最优参数
best_params = space_eval(spaces,best)
print("best_params = ",best_params)


# 5，绘制搜索过程
losses = [x["result"]["loss"] for x in trials.trials]
minlosses = [np.min(losses[0:i+1]) for i in range(len(losses))] 
steps = range(len(losses))

fig,ax = plt.subplots(figsize=(6,3.7),dpi=144)
ax.scatter(x = steps, y = losses, alpha = 0.3)
ax.plot(steps,minlosses,color = "red",axes = ax)
plt.xlabel("step")
plt.ylabel("loss")


#================================================================================
# 三，训练模型
#================================================================================
printlog("step3: training model...")

params.update(best_params)
results = {}
gbm = lgb.train(params,
                lgb_train,
                num_boost_round= boost_round,
                valid_sets=(lgb_valid, lgb_train),
                valid_names=('validate','train'),
                early_stopping_rounds = early_stop_rounds,
                evals_result= results,
                verbose_eval = True)

#================================================================================
# 四，评估模型
#================================================================================
printlog("step4: evaluating model ...")

y_pred_train = gbm.predict(dftrain.drop('label',axis = 1), num_iteration=gbm.best_iteration)
y_pred_test = gbm.predict(dftest.drop('label',axis = 1), num_iteration=gbm.best_iteration)

train_score = f1_score(dftrain['label'],y_pred_train>0.5)
val_score = f1_score(dftest['label'],y_pred_test>0.5)

print('train f1_score: {:.5} '.format(train_score))
print('valid f1_score: {:.5} \n'.format(val_score))

fig2,ax2 = plt.subplots(figsize=(6,3.7),dpi=144)
fig3,ax3 = plt.subplots(figsize=(6,3.7),dpi=144)
lgb.plot_metric(results,ax = ax2)
lgb.plot_importance(gbm,importance_type = "gain",ax=ax3)

#================================================================================
# 五，保存模型
#================================================================================
printlog("step5: saving model ...")


model_dir = "gbm.model"
print("model_dir: %s"%model_dir)
gbm.save_model("gbm.model",num_iteration=gbm.best_iteration)
printlog("task end...")

###
##
#

伙计们，来看看结果吧！

很好，经过自动化调参后，我们的f1_score上升了2个百分点，到了0.98925。?

train f1_score: 1.0 
valid f1_score: 0.98925

本文参与腾讯云自媒体分享计划，分享自微信公众号。

原始发表：2021-07-31，如有侵权请联系 cloudcommunity@tencent.com 删除

自动化

编程算法

本文分享自算法美食屋微信公众号，前往查看

如有侵权，请联系 cloudcommunity@tencent.com 删除。

本文参与腾讯云自媒体分享计划，欢迎热爱写作的你一起参与！

自动化

编程算法

登录后参与评论

0 条评论

热度