Generator functions allow you to declare a function that behaves like an iterator, i.e. it can be used in a for loop.

Simplified Code

The simplification of code is a result of generator function and generator expression support provided by Python.

To illustrate this, we will compare different implementations that implement a function, "firstn", that represents the firstn non-negative integers, wheren is a really big number, and assume (for the sake of the examples in this section) that each integer takes up a lot of space, say 10 megabytes each.

Note: Please note that in real life, integers do not take up that much space, unless they are really, really, really, big integers. For instance you can represent a 309 digit number with 128 bytes (add some overhead, it will still be less than 150 bytes).

First, let us consider the simple example of building a list and returning it.

   1deffirst_n(n):   2'''Build and return a list'''   3num,nums =0, []   4whilenum <n:   5nums.append(num)   6num +=1   7returnnums   8   9  10sum_of_first_n =sum(first_n(1000000))

The code is quite simple and straightforward, but it builds the full list in memory. This is clearly not acceptable in our case, because we cannot afford to keep alln "10 megabyte" integers in memory.

So, we resort to the generator pattern. The following implements generator as an iterable object.

   1# Using the generator pattern (an iterable)   2classfirst_n(object):   3   4   5def__init__(self,n):   6self.n =n   7self.num =0   8   9  10def__iter__(self):  11returnself  12  13  14# Python 3 compatibility  15def__next__(self):  16returnself.next()  17  18  19defnext(self):  20ifself.num <self.n:  21cur,self.num =self.num,self.num+1  22returncur  23raiseStopIteration()  24  25  26sum_of_first_n =sum(first_n(1000000))

This will perform as we expect, but we have the following issues:

there is a lot of boilerplate
the logic has to be expressed in a somewhat convoluted way

Furthermore, this is a pattern that we will use over and over for many similar constructs. Imagine writing all that just to get an iterator.

Python provides generator functions as a convenient shortcut to building iterators. Lets us rewrite the above iterator as a generator function:

   1# a generator that yields items instead of returning a list   2deffirstn(n):   3num =0   4whilenum <n:   5yieldnum   6num +=1   7   8sum_of_first_n =sum(firstn(1000000))

Note that the expression of the number generation logic is clear and natural. It is very similar to the implementation that built a list in memory, but has the memory usage characteristic of the iterator implementation.

Note: the above code is perfectly acceptable for expository purposes, but remember that in Python 2firstn() is equivalent to the built-inxrange() function, and in Python 3range()is an immutable sequence type. The built-ins will always be much faster. SH

Generator expressions provide an additional shortcut to build generators out of expressions similar to that of list comprehensions.

In fact, we can turn a list comprehension into a generator expression by replacing the square brackets ("[ ]") with parentheses. Alternately, we can think of list comprehensions as generator expressions wrapped in a list constructor.

Consider the following example:

   1# list comprehension   2doubles = [2 *nforninrange(50)]   3   4# same as the list comprehension above   5doubles =list(2 *nforninrange(50))

Notice how a list comprehension looks essentially like a generator expression passed to a list constructor.

By allowing generator expressions, we don't have to write a generator function if we do not need the list. If only list comprehensions were available, and we needed to lazily build a set of items to be processed, we will have to write a generator function.

This also means that we can use the same syntax we have been using for list comprehensions to build generators.

Keep in mind that generators are a special type of iterator, and that containers likelist andset are also iterables. The uniform way in which all of these are handled adds greatly to the simplification of code.

Improved Performance

The performance improvement from the use of generators is the result of the lazy (on demand) generation of values, which translates to lower memory usage. Furthermore, we do not need to wait until all the elements have been generated before we start to use them. This is similar to the benefits provided by iterators, but the generator makes building iterators easy.

This can be illustrated by comparing therange andxrange built-ins of Python 2.x.

Bothrange andxrange represent a range of numbers, and have the same function signature, butrange returns alist whilexrange returns a generator (at least in concept; the implementation may differ).

Say, we had to compute the sum of the firstn, say 1,000,000, non-negative numbers.

   1# Note: Python 2.x only   2# using a non-generator   3sum_of_first_n =sum(range(1000000))   4   5# using a generator   6sum_of_first_n =sum(xrange(1000000))

Note that both lines are identical in form, but the one usingrange is much more expensive.

When we userange we build a 1,000,000 element list in memory and then find its sum. This is a waste, considering that we use these 1,000,000 elements just to compute the sum.

This waste becomes more pronounced as the number of elements (ourn) becomes larger, the size of our elements become larger, or both.

On the other hand, when we usexrange, we do not incur the cost of building a 1,000,000 element list in memory. The generator created by xrange will generate each number, whichsum will consume to accumulate the sum.

In the case of the "range" function, using it as an iterable is the dominant use-case, and this is reflected in Python 3.x, which makes therange built-in return a sequence-type object instead of a list.

Note: a generator will provide performance benefits only if we do not intend to use that set of generated values more than once.

Consider the following example:

   1# Note: Python 2.x only   2s =sum(xrange(1000000))   3p =product(xrange(1000000))

Imagine that making a integer is a very expensive process. In the above code, we just performed the same expensive process twice. In cases like this, building a list in memory might be worth it (see example below):

   1# Note: Python 2.x only   2nums =list(xrange(1000000))   3s =sum(nums)   4p =product(nums)

However, a generator might still be the only way, if the storage of these generated objects in memory is not practical, and it might be worth to pay the price of duplicated expensive computations.

Examples

For example, theRangeGenerator can be used to iterate over a large number of values, without creating a massive list (like range would)

   1#the for loop will generate each i (i.e. 1,2,3,4,5, ...), add it to total,  and throw it away   2#before the next i is generated.  This is opposed to iterating through range(...), which creates   3#a potentially massive list and then iterates through it.   4total =0   5foriinirange(1000000):   6total +=i

Generators can be composed. Here we create a generator on the squares of consecutive integers.

   1#square is a generator   2square = (i*iforiinirange(1000000))   3#add the squares   4total =0   5foriinsquare:   6total +=i

Here, we compose a square generator with the takewhile generator, to generate squares less than 100

   1#add squares less than 100   2square = (i*iforiincount())   3bounded_squares =takewhile(lambdax :x<100,square)   4total =0   5foriinbounded_squares:   6total +=i

to be written: Generators made from classes?

Discussion

I once sawMikeOrr demonstrate Before and After examples. But, I forget how they worked.

Can someone demonstrate here?

He did something like: Show how a normal list operation could be written to use generators. Something like:

   1defdouble(L):   2return [x*2forxinL]   3   4eggs =double([1,2,3,4,5])

...he showed how that, or something like that, could be rewritten using iterators, generators.

It's been a while since I've seen it, I may be getting this all wrong.

--LionKimbro 2005-04-02 19:12:19

   1# explicitly write a generator function   2defdouble(L):   3forxinL:   4yieldx*2   5   6# eggs will be a generator   7eggs =double([1,2,3,4,5])   8   9# the above is equivalent to ("generator comprehension"?)  10eggs = (x*2forxin [1,2,3,4,5])  11  12# need to do this if you need a list  13eggs =list(double([1,2,3,4,5]))  14  15# the above is equivalent to (list comprehension)  16eggs = [x*2forxin [1,2,3,4,5]]

For the above example, a generator comprehension or list comprehension is sufficient unless you need to apply that in many places.

Also, a generator function will be cleaner and more clear, if the generated expressions are more complex, involve multiple steps, or depend on additional temporary state.

Consider the following example:

   1defunique(iterable,key=lambdax:x):   2seen =set()   3forelem,ekeyin ((e,key(e))foreiniterable):   4ifekeynotinseen:   5yieldelem   6seen.add(ekey)

Here, the temporary keys collector,seen, is a temporary storage that will just be more clutter in the location where this generator will be used.

Even if we were to use this only once, it is worth writing a function (for the sake of clarity; remember that Python allows nested functions).

Movatterモバイル変換

Simplified Code

Improved Performance

Examples

Links

Discussion