使用 python joblib 访问和更改全局数组

Accessing and altering a global array using python joblib

我正在尝试使用 python 中的 joblib 来加速某些数据处理,但我在尝试找出如何将输出分配为所需格式时遇到了问题。我试图生成一个可能过于简单的代码来显示我遇到的问题:

from joblib import Parallel, delayed
import numpy as np

def main():
    print "Nested loop array assignment:"
    regular()
    print "Parallel nested loop assignment using a single process:"
    par2(1)
    print "Parallel nested loop assignment using multiple process:"
    par2(2)

def regular():
    # Define variables
    a = [0,1,2,3,4]
    b = [0,1,2,3,4]
    # Set array variable to global and define size and shape
    global ab
    ab = np.zeros((2,np.size(a),np.size(b)))

    # Iterate to populate array
    for i in range(0,np.size(a)):
        for j in range(0,np.size(b)):
            func(i,j,a,b)

    # Show array output
    print ab

def par2(process):
    # Define variables
    a2 = [0,1,2,3,4]
    b2 = [0,1,2,3,4]
    # Set array variable to global and define size and shape
    global ab2
    ab2 = np.zeros((2,np.size(a2),np.size(b2)))

    # Parallel process in order to populate array
    Parallel(n_jobs=process)(delayed(func2)(i,j,a2,b2) for i in xrange(0,np.size(a2)) for j in xrange(0,np.size(b2)))

    # Show array output
    print ab2

def func(i,j,a,b):
    # Populate array
    ab[0,i,j] = a[i]+b[j]
    ab[1,i,j] = a[i]*b[j]

def func2(i,j,a2,b2):
    # Populate array
    ab2[0,i,j] = a2[i]+b2[j]
    ab2[1,i,j] = a2[i]*b2[j]

# Run script
main()

其输出如下所示:

Nested loop array assignment:
[[[  0.   1.   2.   3.   4.]
  [  1.   2.   3.   4.   5.]
  [  2.   3.   4.   5.   6.]
  [  3.   4.   5.   6.   7.]
  [  4.   5.   6.   7.   8.]]

 [[  0.   0.   0.   0.   0.]
  [  0.   1.   2.   3.   4.]
  [  0.   2.   4.   6.   8.]
  [  0.   3.   6.   9.  12.]
  [  0.   4.   8.  12.  16.]]]
Parallel nested loop assignment using a single process:
[[[  0.   1.   2.   3.   4.]
  [  1.   2.   3.   4.   5.]
  [  2.   3.   4.   5.   6.]
  [  3.   4.   5.   6.   7.]
  [  4.   5.   6.   7.   8.]]

 [[  0.   0.   0.   0.   0.]
  [  0.   1.   2.   3.   4.]
  [  0.   2.   4.   6.   8.]
  [  0.   3.   6.   9.  12.]
  [  0.   4.   8.  12.  16.]]]
Parallel nested loop assignment using multiple process:
[[[ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]]

 [[ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]
  [ 0.  0.  0.  0.  0.]]]

从 Google 和 Whosebug 搜索函数看来,当使用 joblib 时,全局数组不会在每个子进程之间共享。我不确定这是 joblib 的限制还是有办法解决这个问题?

实际上我的脚本被其他代码包围,这些代码依赖于这个全局数组的最终输出在 (4,x,x ) 格式,其中 x 是可变的(但通常在 100 到几千之间)。这是我目前考虑并行处理的原因,因为 x = 2400.

的整个过程最多可能需要 2 个小时

joblib 的使用不是必需的(但我喜欢命名法和简单性)所以请随意提出简单的替代方法,最好牢记最终数组的要求。我正在使用 python 2.7.3 和 joblib 0.7.1.

我能够使用 numpy 的内存映射解决这个简单示例的问题。 使用 memmap 并遵循 joblib documentation webpage 上的示例后,我仍然遇到问题,但我通过 pip 升级到最新的 joblib 版本 (0.9.3),并且一切运行顺利。这是工作代码:

from joblib import Parallel, delayed
import numpy as np
import os
import tempfile
import shutil

def main():

    print "Nested loop array assignment:"
    regular()

    print "Parallel nested loop assignment using numpy's memmap:"
    par3(4)

def regular():
    # Define variables
    a = [0,1,2,3,4]
    b = [0,1,2,3,4]

    # Set array variable to global and define size and shape
    global ab
    ab = np.zeros((2,np.size(a),np.size(b)))

    # Iterate to populate array
    for i in range(0,np.size(a)):
        for j in range(0,np.size(b)):
            func(i,j,a,b)

    # Show array output
    print ab

def par3(process):

    # Creat a temporary directory and define the array path
    path = tempfile.mkdtemp()
    ab3path = os.path.join(path,'ab3.mmap')

    # Define variables
    a3 = [0,1,2,3,4]
    b3 = [0,1,2,3,4]

    # Create the array using numpy's memmap
    ab3 = np.memmap(ab3path, dtype=float, shape=(2,np.size(a3),np.size(b3)), mode='w+')

    # Parallel process in order to populate array
    Parallel(n_jobs=process)(delayed(func3)(i,a3,b3,ab3) for i in xrange(0,np.size(a3)))

    # Show array output
    print ab3

    # Delete the temporary directory and contents
    try:
        shutil.rmtree(path)
    except:
        print "Couldn't delete folder: "+str(path)

def func(i,j,a,b):
    # Populate array
    ab[0,i,j] = a[i]+b[j]
    ab[1,i,j] = a[i]*b[j]

def func3(i,a3,b3,ab3):
    # Populate array
    for j in range(0,np.size(b3)):
        ab3[0,i,j] = a3[i]+b3[j]
        ab3[1,i,j] = a3[i]*b3[j]

# Run script
main()

给出以下结果:

Nested loop array assignment:
[[[  0.   1.   2.   3.   4.]
  [  1.   2.   3.   4.   5.]
  [  2.   3.   4.   5.   6.]
  [  3.   4.   5.   6.   7.]
  [  4.   5.   6.   7.   8.]]

 [[  0.   0.   0.   0.   0.]
  [  0.   1.   2.   3.   4.]
  [  0.   2.   4.   6.   8.]
  [  0.   3.   6.   9.  12.]
  [  0.   4.   8.  12.  16.]]]
Parallel nested loop assignment using numpy's memmap:
[[[  0.   1.   2.   3.   4.]
  [  1.   2.   3.   4.   5.]
  [  2.   3.   4.   5.   6.]
  [  3.   4.   5.   6.   7.]
  [  4.   5.   6.   7.   8.]]

 [[  0.   0.   0.   0.   0.]
  [  0.   1.   2.   3.   4.]
  [  0.   2.   4.   6.   8.]
  [  0.   3.   6.   9.  12.]
  [  0.   4.   8.  12.  16.]]]

我的一些想法供未来的读者注意:

  • 在小型阵列上,准备并行环境所花费的时间 (通常称为开销)意味着这个运行速度比 简单的 for 循环。
  • 比较更大的数组,例如。将 aa3 设置为 np.arange(0,10000)bb3np.arange(0,1000) 给了 "regular" 方法耗时 12.4 秒,joblib 耗时 7.7 秒 方法。
  • 开销意味着让每个核心执行速度更快 内部 j 循环(参见 func3)。这是有道理的,因为我只是 启动 10,000 个进程而不是启动 10,000,000
    每个过程都需要设置。

我正在使用的 joblib 版本 (0.13.2),实际上允许我访问大共享 DataFrames 而无需太多麻烦。

当然 DataFrames 需要在并行循环开始之前 pre-allocated 并且每个线程必须只访问它的 DataFrame 部分来写入,但它有效。

data  = pd.DataFrame(...)
stats = pd.DataFrame(np.nan, index=np.arange(0, size/step), columns=cols, dtype=np.float64)

Parallel(n_jobs=8, prefer='threads')(
            delayed(_threadsafe_func)(data, stats, i, step, other_params)
            for i in range(0, size, step))

_threadsafe_func 中,可以这样读取或写入 stats DataFrame

index = i/step
print('[' + str(i) + '] Running job with index:', str(int(index)), '/', len(data)/step)
chunk = data[i:i + step]
stats.loc[index, 'mean'] = chunk.mean()    # 'mean' is an existing column already filled with np.nan