Hyperopt是实现超参数优化的python第三方库, 最近发现其可以运用mongo进行并行计算, 稍微研究了一番,记录并分享一下.
Mongo的安装就不说了, 遵循链接内容即可
安装完成后启动mongo, 运行下官方的demo看一下:
import math
from hyperopt import fmin, tpe, hp
from hyperopt.mongoexp import MongoTrials
trials = MongoTrials('mongo://localhost:1234/foo_db/jobs', exp_key='exp1')
best = fmin(math.sin, hp.uniform('x', -2, 2), trials=trials, algo=tpe.suggest, max_evals=10)
以上的代码中, 实例化 MongoTrials 并赋值给trials变量, 其第一个参数是 mongo 进程, 数据库是 'foodb', 'jobs' 表. 'exp_key' 是任务的编号.(如果修改这个参数, 表明是一个新的任务, 会重新运行搜索而不是从数据库中取结果).
实际运行demin的过程中, fmin 会被阻塞. 这是因为 MongoTrials 会将 fmin 作为异步对象, 所以出现新的搜索点(参数组合)时, fmin 不会去评估目标函数而是等待另一个进程替它完成这个工作.
hyperopt-mongo-worker 脚本就是干这个活滴, 新开一个 shell 输入
hyperopt-mongo-worker --mongo=localhost:1234/foo_db --poll-interval=0.1
第一个参数就是 mongo 的地址, 第二个参数是轮询间隔. 由于demo很简单, 我们很快就得到一个最优的 x 值.
但以上的demo太简单了, 我们想将自己编写的模型替换掉 math.sin. 以一个随机森林举例:
import hyperopt.mongoexp
import pandas as pd
import numpy as np
from hyperopt import fmin, tpe, hp, space_eval, pyll, rand, anneal
from hyperopt.mongoexp import MongoTrials
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score, cross_val_predict, train_test_split
def randomforest(args):
class_weight = args['class_weight']
criterion = args['criterion']
min_impurity_split = args['min_impurity_split']
n_estimators = args['n_estimators']
min_samples_leaf = args['min_samples_leaf']
min_samples_split = args['min_samples_split']
estim = RandomForestClassifier(
n_estimators=n_estimators,
class_weight=class_weight,
criterion=criterion,
min_impurity_decrease=min_impurity_split,
min_samples_leaf=min_samples_leaf,
min_samples_split = min_samples_split
)
y_pred = cross_val_predict(estim, train_x, train_y, cv=3)
metric = f1_score(train_y, y_pred)
return -metric
space = {
'class_weight': hp.choice('class_weight', [None, 'balanced']),
'criterion': hp.choice('criterion', ['gini', 'entropy']),
'min_impurity_split': hp.lognormal('min_impurity_split', 1e-10, 1e-4)*1e-7,
'min_samples_leaf': hp.randint('min_samples_leaf', 10)+1,
'min_samples_split': hp.randint('min_samples_split', 10)+1,
'n_estimators': hp.randint('n_estimators', 950)+50
}
if __name__ == '__main__':
trials = MongoTrials('mongo://localhost:1234/foo_db/jobs', exp_key='exp2')
best = fmin(fn=randomforest, space=space, algo=rand.suggest, max_evals=100, trials=trials)
print best
很遗憾有个属性错误, 就是找不到 randomforest 这个模块.
AttributeError: Can't get attribute 'randomforest' on <module '__main__' from ...hyperopt-mongo-worker
google了一下, 有网友给出了一些解决办法, 我们先将 objective function 写到另外的脚本中, 例如:
# hyperopt_model.py
# !-*- coding: utf-8 -*-
from sklearn.metrics import accuracy_score
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score, cross_val_predict, train_test_split
import pandas as pd
df = pd.read_csv('xxxxx.csv', header=0)
y, X = df[df.columns[0]], df[df.columns[1:]]
def randomforest(args):
n_estimators = args['n_estimators']
criterion = args['criterion']
max_features = args['max_features']
min_impurity_split = args['min_impurity_split']
min_samples_leaf = args['min_samples_leaf']
min_samples_split = args['min_samples_split']
class_weight = args['class_weight']
global X, y
clf = RandomForestClassifier(
class_weight=class_weight,
criterion=criterion,
max_features=max_features,
min_samples_leaf=min_samples_leaf,
min_impurity_split=min_impurity_split,
min_samples_split=min_samples_split,
n_estimators=n_estimators,
random_state=1
)
y_pred = cross_val_predict(clf, X, y, cv=3)
metric = accuracy_score(y, y_pred)
return -metric
将这个脚本命名为 hyperopt_model.py 并将其写入环境变量中, 顺便修改下最上面的脚本:
export PYTHONPATH="${PYTHONPATH}:<hyperopt_model.py>"
import pandas as pd
import numpy as np
import hyperopt_model
import hyperopt.mongoexp
from hyperopt import fmin, tpe, hp, space_eval, pyll, rand, anneal
from hyperopt.mongoexp import MongoTrials
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import cross_val_score, cross_val_predict, train_test_split
if __name__ == '__main__':
trials = MongoTrials('mongo://localhost:1234/foo_db/jobs', exp_key='exp2')
best = fmin(fn=hyperopt_model.randomforest, space=hyperopt_model.space, algo=rand.suggest, max_evals=100, trials=trials)
print best
之后再运行 hyperopt-mongo-worker 就ok了, 总体时间消耗大概降低了50% 左右.
我还尝试了用进程管理池管理这两个进程(代码如下), 但是总有一些error没有解决, 如果那位大佬有更好的方法, 烦请告知, 感谢!
# coding: utf-8
import sys
import logging
import hyperopt_model
from multiprocessing import Pool, Process
from hyperopt import fmin, tpe, hp, rand
from hyperopt.mongoexp import MongoTrials
def task1():
logging.basicConfig(stream=sys.stderr, level=logging.INFO)
print 'task1 running'
sys.exit(hyperopt.mongoexp.main_worker())
def task2(msg):
trials = MongoTrials('mongo://localhost:1234/foo_db/jobs', exp_key='exp3')
best = fmin(fn=hyperopt_model.randomforest, space=hyperopt_model.space, algo=rand.suggest, max_evals=100, trials=trials)
print msg
print 'task2 is running'
return best
if __name__ == '__main__':
pool = Pool(processes=4)
p = Process(target=task1)
p.start()
ret = pool.apply_async(task2, args=(1,))
pool.close()
pool.join()
p.join()
print 'processes done, result:'
print ret.get()
### hyperopt ### MongoDB ### 并行计算 ### 自定义超参优化模型
网友评论