Why does a generator expression need a lot of memory?

Question

Problem

Let’s assume that I want to find n**2 for all numbers smaller than 20000000.

General setup for all three variants that I test:

import time, psutil, gc

gc.collect()
mem_before = psutil.virtual_memory()[3]
time1 = time.time()

# (comprehension, generator, function)-code comes here

time2 = time.time()
mem_after =  psutil.virtual_memory()[3]

print "Used Mem = ", (mem_after - mem_before)/(1024**2)  # convert Byte to Megabyte
print "Calculation time = ", time2 - time1

Three options to calculate these numbers:

1. Creating a list of via comprehension:

x = [i**2 for i in range(20000000)]

It is really slow and time consuming:

Used Mem =  1270  # Megabytes
Calculation time =  33.9309999943  # Seconds

2. Creating a generator using '()':

x = (i**2 for i in range(20000000))

It is much faster than option 1, but still uses a lot of memory:

Used Mem =  611 
Calculation time =  0.278000116348

3. Defining a generator function (most efficient):

def f(n):
    i = 0
    while i < n:
        yield i**2
        i += 1
x = f(20000000)

Its consumption:

Used Mem =  0
Calculation time =  0.0

The questions are:

What’s the difference between the first and second solutions? Using () creates a generator, so why does it need a lot of memory?
Is there any built-in function equivalent to my third option?

Asked By: EbraHim

||

Source

Answer 1

As others have pointed out in the comments, range creates a list in Python 2. Hence, it is not the generator per se that uses up the memory, but the range that the generator uses:
```
x = (i**2 for i in range(20000000))  
# builds a 2*10**7 element list, not for the squares , but for the bases

>>> sys.getsizeof(range(100))
872
>>> sys.getsizeof(xrange(100))
40
>>> sys.getsizeof(range(1000))
8720
>>> sys.getsizeof(xrange(1000))
40
>>> sys.getsizeof(range(20000000))
160000072
>>> sys.getsizeof(xrange(20000000))
40
```
This also explains why your second version (the generator expression) uses around half the memory of the first version (the list comprehension) as the first one builds two lists (for the bases and the squares) while the second only builds one list for the bases.
xrange(20000000) thus, greatly improves memory usage as it returns a lazy iterable. This is essentially the built-in memory efficient way to iterate over a range of numbers that mirrors your third version (with the added flexibility of start, stop and step):
```
x = (i**2 for i in xrange(20000000))
```
In Python 3, range is essentially what xrange used to be in Python 2.
However, the Python 3 range object has some nice features that Python 2’s xrange doesn’t have, like O(1) slicing, contains, etc.

Some references:

Answered By: user2390182

Answer 2

1.- The object must be created in memory, so in your second solution, the generator is created but not computed, but still has memory, python probably reserve some memory for its computation to be efficient, we don’t know about the interpreter magic, also notice that range funtion creates the full list from 0 to 200000, so in fact you are still building that list in memory.

2.- You can use itertool.imap:

squares = itertools.imap(lambda x: x**2, xrange(200000))

Answered By: Netwave