如何使用 python 加速随机选择生成？

Question

我用 python 和 numpy 做了一个随机游走生成器。给定一个邻接矩阵，我想从每个节点中抽取随机路径。为此，我目前制作了这个类，它将图中每个节点的邻居节点及其相应的概率作为输入：

import numpy as np

class RandomChoice(object):
    def __init__(self, neighbors_choices, neighbors_prob, depth=50):
        C = len(neighbors_choices)
        self.depth = depth
        self.neighbors_choices = neighbors_choices
        self.neighbors_prob = neighbors_prob
        self.index = np.zeros(C, np.uint32)
        self.choices = list()
        for i in range(C):
            self.choices.append(np.random.choice(self.neighbors_choices[i], size=self.depth, p=self.neighbors_prob[i]))

    def __getitem__(self, arg):
        if self.index[arg] == self.depth:
            self.choices[arg] = np.random.choice(self.neighbors_choices[arg], size=self.depth, p=self.neighbors_prob[arg])
            self.index[arg] = 0
        val = self.choices[arg][self.index[arg]]
        self.index[arg] += 1
        return val

我是这样使用它的：

# Example with a 3x3 matrix
#  1 2 1
#  3 0 1
#  0 1 0

number_of_walks_per_node = 5
number_of_nodes = 3
length_of_walks = 10

choices = [
            [0, 1, 2],
            [0, 2],
            [1]
          ]
probs =   [
            [0.25, 0.5, 0.25],
            [0.75, 0.25],
            [1]
          ]
randomChoice = RandomChoice(choices, probs, depth=50)
for i in range(number_of_walks_per_node):
    for starting_node in range(number_of_nodes):
        walker_positions = [starting_node]
        for j in range(length_of_walks):
            walker_positions.append( randomChoice[walker_positions[j]])
        print(walker_positions)

这里的想法是利用 numpy.random.choices 对 RAM 中的一些空间的向量效率。但是这个功能仍然是这段代码的瓶颈。我认为 numpy.random.choices 花时间检查概率总和是否为 1，并且每个概率都大于 0。你知道如何加快这段代码的速度吗？

Answer 1

我更新了您的代码以使其运行（见下文）。这是我得到的分析：

看起来 getitem 上的函数开销与其他所有内容相比非常大，但我预计这部分是因为玩具图。

如果通常是这种情况，可以重构您的代码，使您不那么频繁地调用 getitem。相反，您可以将代码从 getitem 移动到脚本中的一组嵌套循环中，就像这样（但它会很丑）...

for i in range(number_of_walks_per_node):
    for starting_node in range(number_of_nodes):
        walkers_positions = [start_node]
        for j in range(length_of_walks):
            if randomChoice.index[walkers_positions[j]] == randomChoice.depth:
                randomChoice.choices[walkers_positions[j]] = np.random.choice(randomChoice.neighbors_choices[walkers_positions[j]], size=randomChoice.depth,
                                                     p=randomChoice.neighbors_prob[walkers_positions[j]])
                randomChoice.index[walkers_positions[j]] = 0
            val = randomChoice.choices[walkers_positions[j]][randomChoice.index[walkers_positions[j]]]
            randomChoice.index[walkers_positions[j]] += 1
            walkers_positions.append( val )

您的代码的工作版本以供参考...

import numpy as np

class RandomChoice(object):
    def __init__(self, neighbors_choices, neighbors_prob, depth=50):
        C = len(neighbors_choices)
        self.depth = depth
        self.neighbors_choices = neighbors_choices
        self.neighbors_prob = neighbors_prob
        self.index = np.zeros(C, np.uint32)
        self.choices = list()
        for i in range(C):
            self.choices.append(np.random.choice(self.neighbors_choices[i], size=self.depth, p=self.neighbors_prob[i]))

    def __getitem__(self, arg):
        if self.index[arg] == self.depth:
            self.choices[arg] = np.random.choice(self.neighbors_choices[arg], size=self.depth, p=self.neighbors_prob[arg])
            self.index[arg] = 0
        val = self.choices[arg][self.index[arg]]
        self.index[arg] += 1
        return val

# Example with a 3x3 matrix
#  1 3 1
#  3 0 1
#  0 1 0

choices = [
            [0, 1, 2],
            [0, 2],
            [1]
          ]
probs =   [
            [0.25, 0.5, 0.25],
            [0.75, 0.25],
            [1]
          ]

number_of_walks_per_node = 10
number_of_nodes = 3
start_node = 0
length_of_walks = 5000
randomChoice = RandomChoice(choices, probs, depth=50)
for i in range(number_of_walks_per_node):
    for starting_node in range(number_of_nodes):
        walkers_positions = [start_node]
        for j in range(length_of_walks):
            walkers_positions.append( randomChoice[walkers_positions[j]] )