How to use random.shuffle() on a generator? python

问题:

How do I use random.shuffle() on a generator without initializing a list from the generator? Is that even possible? if not, how else should I use random.shuffle() on my list?

>>> import random
>>> random.seed(2)
>>> x = [1,2,3,4,5,6,7,8,9]
>>> def yielding(ls):
...     for i in ls:
...             yield i
... 
>>> for i in random.shuffle(yielding(x)):
...     print i
... 
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/lib/python2.7/random.py", line 287, in shuffle
    for i in reversed(xrange(1, len(x))):
TypeError: object of type 'generator' has no len()

Note: random.seed() was designed such that it returns the same output after each script run?

回答1:

In order to shuffle the sequence uniformly, random.shuffle() needs to know how long the input is. A generator cannot provide this; you have to materialize it into a list:

lst = list(yielding(x))
random.shuffle(lst)
for i in lst:
    print i

You could, instead, use sorted() with random.random() as the key:

for i in sorted(yielding(x), key=lambda k: random.random()):
    print i

but since this also produces a list, there is little point in going this route.

Demo:

>>> import random
>>> x = [1,2,3,4,5,6,7,8,9]
>>> sorted(iter(x), key=lambda k: random.random())
[9, 7, 3, 2, 5, 4, 6, 1, 8]

回答2:

It's not possible to randomize the yield of a generator without temporarily saving all the elements somewhere. Luckily, this is pretty easy in Python:

tmp = list(yielding(x))
random.shuffle(tmp)
for i in tmp:
    print i

Note the call to list() which will read all items and put them into a list.

If you don't want to or can't store all elements, you will need to change the generator to yield in a random order.

回答3:

I needed to find a solution to this problem so I could get expensive to compute elements in a shuffled order, without wasting computation by generating values. This is what I have come up with for your example. It involves making another function to index the first array.

You will need numpy installed

pip install numpy

The Code:

import numpy as np
x = [1, 2, 3, 4, 5, 6, 7, 8, 9]

def shuffle_generator(lst):
    return (lst[idx] for idx in np.random.permutation(len(lst)))

def yielding(ls):
    for i in ls:
        yield i

# for i in random.shuffle(yielding(x)):
#    print i

for i in yielding(shuffle_generator(x)):
    print(i)

回答4:

Depending on the case, if you know how much data you have ahead of time, you can index the data and compute/read from it based on a shuffled index. This amounts to: 'don't use a generator for this problem', and without specific use-cases it's hard to come up with a general method.

Alternatively... If you need to use the generator...

it depends on 'how shuffled' you want the data. Of course, like folks have pointed out, generators don't have a length, so you need to at some point evaluate the generator, which could be expensive. If you don't need perfect randomness, you can introduce a shuffle buffer:

from itertools import islice

import numpy as np


def shuffle(generator, buffer_size):
    while True:
        buffer = list(islice(generator, buffer_size))
        if len(buffer) == 0:
            break
        np.random.shuffle(buffer)
        for item in buffer:
            yield item


shuffled_generator = shuffle(my_generator, 256)

This will shuffle data in chunks of buffer_size, so you can avoid memory issues if that is your limiting factor. Of course, this is not a truly random shuffle, so it shouldn't be used on something that's sorted, but if you just need to add some randomness to your data this may be a good solution.