详解pandas apply 并行处理的几种方法_Python

1. pandarallel (pip install )

对于一个带有Pandas DataFrame df的简单用例和一个应用func的函数，只需用parallel_apply替换经典的apply。

									from pandarallel import pandarallel

									# Initialization

									pandarallel.initialize()

									# Standard pandas apply

									df.apply(func)

									# Parallel apply

									df.parallel_apply(func)

注意，如果不想并行化计算，仍然可以使用经典的apply方法。

另外可以通过在initialize函数中传递progress_bar=True来显示每个工作CPU的一个进度条。

2. joblib (pip install )

https://pypi.python.org/pypi/joblib

				?

									# Embarrassingly parallel helper: to make it easy to write readable parallel code and debug it quickly

									from math import sqrt

									from joblib import Parallel, delayed

									def test():

									  start = time.time()

									  result1 = Parallel(n_jobs=1)(delayed(sqrt)(i**2) for i in range(10000))

									  end = time.time()

									  print(end-start)

									  result2 = Parallel(n_jobs=8)(delayed(sqrt)(i**2) for i in range(10000))

									  end2 = time.time()

									  print(end2-end)

-------输出结果----------

0.4434356689453125
0.6346755027770996

3. multiprocessing

				?

									import multiprocessing as mp

									with mp.Pool(mp.cpu_count()) as pool:

									  df['newcol'] = pool.map(f, df['col'])

									multiprocessing.cpu_count()

返回系统的CPU数量。

该数量不同于当前进程可以使用的CPU数量。可用的CPU数量可以由 len(os.sched_getaffinity(0)) 方法获得。

可能引发 NotImplementedError 。

参见os.cpu_count()

4. 几种方法性能比较

（1）代码

				?

									import sys

									import time

									import pandas as pd

									import multiprocessing as mp

									from joblib import Parallel, delayed

									from pandarallel import pandarallel

									from tqdm import tqdm, tqdm_notebook

									def get_url_len(url):

									  url_list = url.split(".")

									  time.sleep(0.01) # 休眠0.01秒

									  return len(url_list)

									def test1(data):

									  """

									  不进行任何优化

									  """

									  start = time.time()

									  data['len'] = data['url'].apply(get_url_len)

									  end = time.time()

									  cost_time = end - start

									  res = sum(data['len'])

									  print("res:{}, cost time:{}".format(res, cost_time))

									def test_mp(data):

									  """

									  采用mp优化

									  """

									  start = time.time()

									  with mp.Pool(mp.cpu_count()) as pool:

									    data['len'] = pool.map(get_url_len, data['url'])

									  end = time.time()

									  cost_time = end - start

									  res = sum(data['len'])

									  print("test_mp \t res:{}, cost time:{}".format(res, cost_time))

									def test_pandarallel(data):

									  """

									  采用pandarallel优化

									  """

									  start = time.time()

									  pandarallel.initialize()

									  data['len'] = data['url'].parallel_apply(get_url_len)

									  end = time.time()

									  cost_time = end - start

									  res = sum(data['len'])

									  print("test_pandarallel \t res:{}, cost time:{}".format(res, cost_time))

									def test_delayed(data):

									  """

									  采用delayed优化

									  """

									  def key_func(subset):

									    subset["len"] = subset["url"].apply(get_url_len)

									    return subset

									  start = time.time()

									  data_grouped = data.groupby(data.index)

									  # data_grouped 是一个可迭代的对象，那么就可以使用 tqdm 来可视化进度条

									  results = Parallel(n_jobs=8)(delayed(key_func)(group) for name, group in tqdm(data_grouped))

									  data = pd.concat(results)

									  end = time.time()

									  cost_time = end - start

									  res = sum(data['len'])

									  print("test_delayed \t res:{}, cost time:{}".format(res, cost_time))

									if __name__ == '__main__':

									  columns = ['title', 'url', 'pub_old', 'pub_new']

									  temp = pd.read_csv("./input.csv", names=columns, nrows=10000)

									  data = temp

									  """

									  for i in range(99):

									    data = data.append(temp)

									  """

									  print(len(data))

									  """

									  test1(data)

									  test_mp(data)

									  test_pandarallel(data)

									  """

									  test_delayed(data)

(2) 结果输出

1k
res:4338, cost time:0.0018074512481689453
test_mp res:4338, cost time:0.2626469135284424
test_pandarallel res:4338, cost time:0.3467681407928467

1w
res:42936, cost time:0.008773326873779297
test_mp res:42936, cost time:0.26111721992492676
test_pandarallel res:42936, cost time:0.33237743377685547

10w
res:426742, cost time:0.07944369316101074
test_mp res:426742, cost time:0.294996976852417
test_pandarallel res:426742, cost time:0.39208269119262695

100w
res:4267420, cost time:0.8074917793273926
test_mp res:4267420, cost time:0.9741342067718506
test_pandarallel res:4267420, cost time:0.6779992580413818

1000w
res:42674200, cost time:8.027287006378174
test_mp res:42674200, cost time:7.751036882400513
test_pandarallel res:42674200, cost time:4.404983282089233

在get_url_len函数里加个sleep语句（模拟复杂逻辑），数据量为1k，运行结果如下：

1k
res:4338, cost time:10.054503679275513
test_mp res:4338, cost time:0.35697126388549805
test_pandarallel res:4338, cost time:0.43415403366088867
test_delayed res:4338, cost time:2.294757843017578

5. 小结

（1）如果数据量比较少，并行处理比单次执行效率更慢；

（2）如果apply的函数逻辑简单，并行处理比单次执行效率更慢。

6. 问题及解决方法

（1）ImportError: This platform lacks a functioning sem_open implementation, therefore, the required synchronization primitives needed will not function, see issue 3770.

https://www.jianshu.com/p/0be1b4b27bde

（2）Linux查看物理CPU个数、核数、逻辑CPU个数

https://lover.blog.csdn.net/article/details/113951192

(3) 进度条的使用

http://www.zzvips.com/article/190442.html

到此这篇关于详解pandas apply 并行处理的几种方法的文章就介绍到这了,更多相关pandas apply 并行处理内容请搜索服务器之家以前的文章或继续浏览下面的相关文章希望大家以后多多支持服务器之家！

原文链接：https://blog.csdn.net/jingyi130705008/article/details/113949730