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

https://www.jb51.net/article/206219.htm

标签:
pandas,apply,并行,pandas,apply

免责声明:本站文章均来自网站采集或用户投稿,网站不提供任何软件下载或自行开发的软件! 如有用户或公司发现本站内容信息存在侵权行为,请邮件告知! 858582#qq.com
评论“详解pandas apply 并行处理的几种方法”
暂无“详解pandas apply 并行处理的几种方法”评论...

RTX 5090要首发 性能要翻倍!三星展示GDDR7显存

三星在GTC上展示了专为下一代游戏GPU设计的GDDR7内存。

首次推出的GDDR7内存模块密度为16GB,每个模块容量为2GB。其速度预设为32 Gbps(PAM3),但也可以降至28 Gbps,以提高产量和初始阶段的整体性能和成本效益。

据三星表示,GDDR7内存的能效将提高20%,同时工作电压仅为1.1V,低于标准的1.2V。通过采用更新的封装材料和优化的电路设计,使得在高速运行时的发热量降低,GDDR7的热阻比GDDR6降低了70%。