June 13, 2018

Python Generators

Python generators are a powerful, but misunderstood tool. They're often treated as too difficult a concept for beginning programmers to learn — creating the illusion that beginners should hold off on learning generators until they are ready. I think this assessment is unfair, and that you can use generators sooner than you think. In this tutorial, we'll cover:

  • The basic terminology needed to understand generators
  • What a generator is
  • How to create your own generators
  • How to use a generator and generator methods
  • When to use a generator


To get the most out of this tutorial, you should be familiar with the following concepts:

  • Basic Python data structures

    • What a list is
    • What a dictionary is
  • Functions

    • What a function is
    • How to create and use functions
  • List Comprehensions

    • What a list comprehension is
    • How to create a simple list comprehension

Basic terms to know

Iteration and iterables

Iteration is the repetition of some kind of process over and over again. Python's

for loop gives us an easy way to iterate over various objects. Often, you'll iterate over a list, but we can also iterate over other Python objects such as strings and dictionaries.

# Iterating over a list
ez_list = [1, 2, 3]
for i in ez_list:
>>> 1
>>> 2
>>> 3

# Iterating over a string
ez_string = "Generators"
for s in ez_string:
>>> "G"
>>> "e"
>>> "r"
>>> "s"

# Iterating over a dictionary
ez_dict = {1 : "First, 2 : "Second"}
for key, value in ez_dict.items():
    print(k, v)
>>> 1 "First"
>>> 2 "Second"

In each of the above examples, the

for loop iterates over the sequence we give it. The code above used a list, string, and dictionary, but you can iterate over tuples and sets as well. In each loop above, we print each of the items in the sequence in the order they appear. For example, you can confirm that the order of the ez_list is replicated in the order that its items are printed out. For Loop Anatomy We refer to any object that can support iteration as an iterable.

What defines an iterable?

Iterables support something called the

Iterator Protocol. The technical definition for the Iterator Protocol is out of the scope of this article, but it can be thought of as a set of requirements to be used for a for loop. That is to say: lists, strings and dictionaries all follow the Iterator Protocol, therefore we can use them in for loops. Conversely, objects that do not follow the protocol cannot be used in a for loop. One example of an object that does not follow the protocol is an integer. If we try to give an integer to a for loop, Python will throw an error.

number = 12345
for n in number:
>>> TypeError: 'int' object is not iterable

An integer is just a singular number, not a sequence. You may argue that the "first" number in

number is 1, but it is not the same as the first item in a sequence. It doesn't make sense to ask "What's after 1?" from number since Python only understands integers as a single entities. Therefore, one of the requirements to be an iterable is to be able to describe to the for loop what the next item to perform the operation on is. For example, lists tell the for loop that the next item to iterate on is in the index+1 from the current one (1 comes after 0). Consequently, an iterable must also signal to a for loop when to stop iterating. This signal usually comes when we arrive at the end of a sequence (i.e. the end of a list or string). We will explore the specific functions that make something iterable later in this article, the important thing to know is that iterables describe how a for loop should traverse its contents. Generators are iterables themselves. As you'll see later, for loops are one of the main ways we use a generator, so they must be able to support iteration. We'll delve into how we can create our own generators in the next secton.

Key takeaways: basic terms to know

  • Iteration is the idea of repeating some process over a sequence of items. In Python, iteration is usually related to the for loop.
  • An iterable is an object that supports iteration.
  • To be an iterable, it must describe to a for loop two things:

    1. What item comes next in the iteration.
    2. When should the loop stop iteration.
  • Generators are iterables.

A data-based approach

To truly explore generators, we'll use the

Brewer's Friend Beer Recipes data set from Kaggle. You can find the data set here, if you'd like to follow along on your own computer. The data contains important beer characteristics from brewers around the world, including style of beer, alcohol by volume (ABV), and amount of beer produced. For the purposes of this article, let's say that we are interested in brewing our own beer. Perhaps we want to sell our beer, so we would like to see what others have done to inform our brewing choices and produce more popular beer styles. Author's Note: The "Name" column in the original data set contains some messy values that interfere with our analysis. You can find a cleaned version that will serve our purposes here.

Generators and you

If you've never encountered a generator before, the most common real-life example of a generator is a backup generator, which creates —

generates — electricity for your house or office. Conceptually, Python generators generate values one at a time from a given sequence, instead of giving the entirety of the sequence at once. This one-at-a-time fashion of generators is what makes them so compatible with for loops. If this sounds confusing, don't worry too much. As we explain how to create generators, it will become more clear. There are two ways to create a generator. They differ in their syntax, but the end result is still a generator. We'll teach these concepts by covering their syntax and comparing them to a similar, but non-generator equivalent.

  1. A generator function versus a regular function
  2. A generator expression versus a list comprehension

The generator function

A generator function is just like a regular function but with a key difference: the

yield keyword replaces return.

# Regular function
def function_a():
    return "a"

# Generator function
def generator_a():
    yield "a"

The two functions above perform exactly same action (returning/yielding the same string). However, if you try to inspect the generator function, it won't match what the regular function shows.

>>> "a"


>>> <generator object a at 0x000001565469DA98>

Calling a regular function tells Python to go back to where the function is located in our code, perform the code within the block, and return the result. In order to get the generator function to yield its values, you need to pass it into the

next() function. next() is a special function that asks, "What's the next item in the iteration?" In fact, next() is the precise function that is called when you run a for loop! Lists, dictionaries, strings, and the like all implement next(), so this is why you can incorporate them into loops in the first place.

# Asking the generator what the next item is
>>> "a"

# Do not do this

Notice that we have to pass in generator function with the parentheses since the function itself is the generator. Providing only the function name will throw an error since you're trying to give

next() a function name. As expected, the generator function will yield "a" once we invoke the next() function. This example is not fully representative of what a generator is useful for. Remember that generators produce a stream of values, so yielding a single value doesn't really qualify as a stream. To do this, we can actually put in multiple yield statements into a generator function. These yield statements form the sequence that the generator will output. We'll create a generator and bind it to a varible mg. Then, if we keep passing mg into next(), we'll get to the next yield. If we keep going past, we'll be given a StopIteration error to tell us that the generator has no more values to give. The StopIteration error is actually how a for loop knows when to stop iterating.

def multi_generate():
    yield "a"
    yield "b"
    yield "c"

mg = multi_generate()
>>> "a"

>>> "b"

>>> "c"

>>> StopIteration: 


multi_generate to mg is a crucial step in using a generator function. Binding a generator to mg allows us to create a single instance of a generator we can refer back to. We can continue passing mg into next() and get those other yield statements. Observe what happens if we just keep trying to pass in multi_generate itself.

>>> "a"
>>> "a"

It's easy to think of generators as a machine that waits for one command and one command only:

next(). Once you call next() on the generator, it will dispense the next value in the sequence it is holding. Otherwise, you can't do much else with a generator. The image below represents our generator as a simple machine. Generators 1 We continue to get the result of the first yield statement. The reason behind this is subtle. When we pass the generator function itself into next(), Python assumes you are passing a new instance of multi_generate into it, so it will always give you the first yield result. By binding the generator to a variable, Python knows you are trying to act on the same thing when you pass it into next(). We've noted that as we keep passing in mg into next, we get the other yield results. This is possible only if the generator somehow remembers what it last did. This memory is what distinguishes generator functions from regular functions! Once you use a function, it's a one-and-done deal. Once you return the value from the function. A generator will keep yielding values until its out. This brings us to another important property of generators. Once we've finished iterating through them, we can't use them anymore. Once we got through all three yield values in mg, it can't provide anything to us anymore. We'd have to store another instance of the multi_generate generator to begin asking next() statements of it again. Our data still hasn't been read in yet, so let's do that with a generator function. The data is called recipeData.csv, and its contained in a CSV file. We'll use the open() function to enable us to read it, and we'll start using next() function to read what the first few lines of the CSV are.

# Creating a generator that will generate the data row by row
def beerDataGenerator():
    file = "recipeData.csv"
    for row in open(file, encoding="ISO-8859-1"):
        yield row

We'll slowly dissect the above code:

  • We've designated dataGenerator as our generator function that will dispense our CSV file row by row. The function includes the name of the file in file, and this enables us to use the open() function to be able to read it.
  • While we've discussed that Python objects like lists and dictionaries can be iterated over, we can also iterate over files that we open() as well.
  • The encoding tells Python what kinds of characters it should expect to see; ISO-8859-1 specifically refers to Latin-1.
  • The for loop will start with the first row in the CSV file, yield that row, and then save its current place in reading the file until the generator function is called again.

If you're following along with the data on your own computer, you'll need to replace

file with the exact path on your computer to where the file is located. This will enable Python to find it when you want to open() it.

# Remember to store an instance of the generator so we can refer back to it
beer = beerDataGenerator()
>>> 'BeerID,Name,URL,Style,StyleID,Size(L),OG,FG,ABV,IBU,Color,BoilSize,BoilTime,BoilGravity,Efficiency,MashThickness,SugarScale,BrewMethod,PitchRate,PrimaryTemp,PrimingMethod,PrimingAmount,UserId\n'next(beer

>>> '1,Vanilla Cream Ale,/homebrew/recipe/view/1633/vanilla-cream-ale,Cream Ale,45,21.77,1.055,1.013,5.48,17.65,4.83,28.39,75,1.038,70,N/A,Specific Gravity,All Grain,N/A,17.78,corn sugar,4.5 oz,116\n'

Once we've created a

beerDataGenerator in beer, we can start passing it into next() to look at the data itself. As the CSV file suggests, the columns are separated by commas. Furthermore, each row ends with an \n, which indicates a line break. We found that the first item in recipeData.csv to is a list of column names and the first row to describe a delicious Vanilla Cream Ale.

A self-imposed restriction

You may be asking, "We can store the data in a list comprehension! Why jump through an extra hoop and use a generator?" As a programmer, you may encounter

Big Data. This is a somewhat nebulous term, and so we won't delve into the various Big Data definitions here. Suffice to say that any Big Data file is too big to assign to a variable. Our data file doesn't qualify as Big Data, but we can still learn a lot by imposing a restriction on ourselves to recreate this conundrum. We'll assume for now that our beer data is so large in size that we are incapable of storing all of the data in a list of lists. With the normal route of reading in data blocked off, we are forced to reconsider our options. This is where generators come in. We'll explain later precisely why generators work here, but until then we can rest assured that our generator function will enable us to read the data in the first place, albeit not all at once. Along with generator functions, we can also create generators using generator expressions.

The generator expression

Early, we compared our generator function to a regular function since they have many similar aspects. For generation expressions, we'll use

list comprehensions.

lc_example = [n**2 for n in [1, 2, 3, 4, 5]]

genex_example = (n**2 for n in [1, 2, 3, 4, 5])

lc_example is our list comprehension, while genex_example is our generator expression that performs almost the same task. Take note that the only difference between the two is that the generator expression is surrounded by parentheses, rather than brackets. If we either of these iterators in a for loop, they will produce the same result and will be indistinguishable. However, if we try to inspect these variables in our interpreter, they produce different results.

>>> [1, 4, 9, 16, 25]

>>> <generator object <genexpr> at 0x00000156547B4FC0>

This result is similar to what we saw when we tried to look at a regular function and a generator function. Python also recognizes that

genex_example is a generator in generator expression form (). As lc_example is a list, we can perform all of the operations that they support: indexing, slicing, mutation, etc. We cannot do this with the generator expression. Generators are specialized as an easy to produce an output one-at-a-time, so they do not support these operations. However, like list comprehensions we can implement logic within generator expressions to form a filter if we needed it.

genex_example2 = (n**2 for n in [1, 2, 3, 4, 5] if n >= 3)
>>> 9

Effectively, there is no difference in how we will use a generator function or generator expression. Once we have our generator expression, we can call

next() on it to start getting the values it will produce. Once we go through all of the values that the generator expression can produce, we cannot use it anymore. This contrasts against a list comprehension, which we can reuse as much as we want.

>>> 1

# Repeat until we reach the end...

>>> 25
>>> StopIteration:

The idea that we can only use generators

once is tied to the idea of their consumption. Recall that when we iterate over some iterator, we perform some operation on each of the values within. We then move on with our analysis using these processed values, meaning that typically we may not need the original iterator. Generators fit perfectly into this need, allowing us to form an iterator that we can use once and then not have to worry about it taking up space after we use it (in a for loop, for example). We talked about next() as the way to get the values from the generators, but its often better to use generators in for loops. Using next() forces us to have to deal with the StopIteration ourselves, but the for loop uses this to know when to stop!

# Using a for loop to consume a generator is better than using next()
for ge in genex_example:
>>> 1
>>> 4
>>> 9
>>> 16
>>> 25

One distinction that generator expressions have over functions is their succinctness. Generator functions take up multiple lines, whereas we can fit generator expressions in one line. Multiple lines are not bad in and of itself, but it opens up functions to greater complexity that may introduce bugs later on. We'll rewrite our generator function as a one-line expression that read in our beer data. This conciseness that will come in handy later in the article.

beer_data = "recipeData.csv"

# This one line perfoms the same action as beerDataGenerator()!

lines =  (line for line in open(beer_data, encoding="ISO-8859-1"))

Key takeaways: generators and you

  • Generators produce values one-at-a-time as opposed to giving them all at once.
  • There are two ways to create generators: generator functions and generator expressions.
  • Generator functions yield, regular functions return.
  • Generator expressions need (), list comprehensions use [].
  • You can only use a generator once.
  • There are two ways to get values from generators: the next() function and a for loop. The for loop is often the preferred method.
  • We can use generators to read files and give us one line at a time.

Generators: motivation and uses

Earlier, we discussed imposing a restriction on ourselves that forced us to use a generator to read our data instead of reading it into a list of lists. We cited the problem of Big Data and an our inability to store it all in one variable. While calling it a Big Data problem is still correct, you may also call it

a memory problem. Let's say that you have an older laptop with about 4GB of RAM, random access memory. The true size of our beer data set is only about 3MB, but suppose that we asked everyone around the globe to give us their recipes, resulting in a data set around 3GB. If we were to read the entirety of our data set into a variable, it would take up a bit more than 3GB of RAM! This would leave us with little room for other operations, much less other variables of similar size. Storing our data in a list of lists would take up so much memory that any analyses we do would take excruciatingly long to do.

Laziness and generators

We know now that generators produce a single value from a defined sequence, but

only when we ask next() or within a for loop. We call this lazy evaluation. Generators are lazy because they only give us a value when we ask for it. The flipside here is that only that single value takes up memory. The ultimate result is that generators are incredibly memory efficient, which makes it a perfect candidate for reading and using Big Data files. Once we ask for the next value of a generator, the old value is discarded. Once we go through the entire generator, it is also discarded from memory as well.

Generators feeding generators

We currently haven't learned anything from the beer data. All we've done so far is to take the original CSV file and create a generator that will yield each line in the CSV, one at a time in the form of a string. Unless we'd like to do some crazy string manipulation, we'll need to think of a way to get our data into a readable, useable form. Below is a representation of what our code currently does: a simple read from file and output of a single line from the file.

Generators 2 Generators come to the rescue again here! So far in the article, we've been passing in other structures, specifically iterators, to the generators to indicate what sequence we'd like to generate from. However, generators are iterators themselves too — why don't we create another generator that takes the output another generator? Our lines generator outputs the line in its entirety, so we'll make a second generator that does some formatting for us.

beer_data = "recipeData.csv"
lines = (line for line in open(beer_data, encoding="ISO-8859-1"))
lists = (l.split(",") for l in lines)

The end result of our generators is a stream of lists, each containing the data within a row of the CSV. If we iterate through

lists, we'll be able to easily access the data elements within and perform the analyses we need! We've effectively made a pipeline for our data set, starting from the raw data set and sending it through 2 generators to get it into a familiar form. Remember that generators aren't lists themselves, they merely generate a single element of a sequence and only take up the amount that element needs. By piping generators together, we've created a quick, easy-to-read way for us to read data that would be inaccessible through normal means. There's some real power to this approach, and its significance can't be understated. We didn't need to create any temporary lists to hold intermediate values as we processed them. With the additional generator in the pipeline, our code might look like this: Generators 3 In this pipeline, each generator is put in charge of a single operation that will eventually be applied to all rows of the data set. Although having each list is good, there's still some small issues that need to be addressed before we can do any meaningul analyses. First, we'd like to take the column names since they aren't data and then turn them into a dictionary that would make any further code easier to read. Note: if you're running this code on your own machine, you must remember that you can only use generators once. If you use the generator in a for loop to view the output, you'll need to run the data and the whole pipeline again. Thankfully, the generators run fast here.

beer_data = "recipeData.csv"
lines = (line for line in open(beer_data, encoding="ISO-8859-1"))
lists = (l.split(",") for l in lines)

# Take the column names out of the generator and store them, leaving only data
columns = next(lists)

# Take these columns and use them to create an informative dictionary
beerdicts = (dict(zip(columns, data)) for data in lists)


beerdicts does some simple formatting, which gives our pipeline even more power! Generators 4 This is a great place to start inquiring our data about our future beer brewing choices. Now that we have our generator pipeline in place, we can start consuming the data produced by the generators and create some insights. We usually consume generators using for loops, so we'll use one to figure out what the most popular type of homebrewed beer is.

beer_counts = {}
for bd in beerdicts:
    if bd["Style"] not in beer_counts:
        beer_counts[bd["Style"]] = 1
        beer_counts[bd["Style"]] += 1

most_popular = 0
most_popular_type = None
for beer, count in beer_counts.items():
    if count > most_popular:
        most_popular = count
        most_popular_type = beer

>>> "American IPA"

This operation is ubiquitous in data wrangling and processing, and you've probably seen it before. The only new thing here is that instead of referring back to a list of lists containing our data, we rely on dictionaries that are produced by our generators. With generators, we are able to make the same inquires we'd want from any Big Data set as we would a regular-sized one. We now know that American IPAs are the most popular homebrewed beer in the data set, and we know how many entries they have in the data. We can try figuring out how strong our beer should be. This data is contained in the "ABV" (

Alcohol By Volume) key. Since we are working with dictionaries as the output of our generator stream, why don't we add another generator to hone in on the exact values we want to output.

abv = (float(bd["ABV"]) for bd in beerdicts if bd["Style"] == "American IPA")

# Get the average ABV for an American IPAsum(abv)/most_popular
>>> 6.44429396984925

This last generator forms the last of our pipeline, as visualized below:

Generators 5 We should take special note of our use of sum() with the abv generator. It is not immediately intuitive that sum() will sum up all of the ABV values that it receives. You may think of sum() as reducing the whole output of the generator into one value. By dividing this sum by the number of American IPA entires there are, we got the average. Our data suggests that your average American IPA is about 6.4% alcohol by volume! Our last generator abv takes the dictionaries that are output by beerdicts and outputs the ABV key, but only if the beer is an American IPA. Filters on our generator expression form a powerful tool in our pipeline. If we think of each successive generator as a modular component, we can then swap out generators for others that may have a more desirable functionality. If we wanted to change what kind of beer we wanted to investigate or look at another beer characteristic, the only thing we need to change is the generator operation. The picture below expresses the different parts of the generator pipeline approach. It consists of some raw data you want to process, the pipeline that does the actual processing, and the final consumption of the output of this pipeline. Following this pattern will enable you to reenact what we've done with the beer data. Generator Pipeline If you're used to the workflow of using a list of lists and leveraging all the list methods to do your analyses, this new approach to data wrangling might be strange. However, the data pipeline is a powerful concept that can be immediately incorporated into your code and you should try it.

Infinite generation

Let's say that it's been a few years since you first made that analysis on the beer data set. You used the insights from your analyses to create a successful American IPA, and you only have generators to thank. You produce thousands of bottles of your beer each day, and you need a way to analyze the quality of each batch. You go back to your old laptop to fire up the Python interpreter. You begin to write up a function to calculate and check the various qualities of your beer batches as you make them, but then you stop. You don't know how much beer you're going to make in advance. Analyzing the batches by day could offer discrete points for us to look at, but what if we want continuous data? Regular functions would not work here! They expect some arguments and will always return a discrete object. You can't conceivably give a regular function a stream of data and return a continuous stream of values. But you can with generators! Generators are well suited to this type of task. We've discussed how generators will yield values one-at-a-time until it's told to stop. If we never give a generator a stopping signal, it will happily generate these values ad infinitum. Take the example below.

function alwaysBeer():
    while True:
        yield "Beer"


while loop will always be true, so the generator function will always be yielding beer. We haven't done anything wrong, this is completely valid code (although not one you'd actually want to implement)! How would this "infinite" stream of beer fit into our timeline? In our original look at the beer data set, the CSV was originally a set amount of lines. If you could automate your brewing process to output this data to a similar CSV and continuously update it, all you need to do to run your analyses would be to run the data through the generators again! You could conceivably create a generator in your pipeline to catch any batches that don't meet your expectations and flag them in real time! Unfortunately, we don't have said data, but this thought experiment should offer another compelling use case for the Python generator. With generators, you could even tackle infinity (in some cases).

Key takeaways: motivation and uses behind generators

  • Generators are memory efficient since they only require memory for the one value they yield.
  • Generators are lazy: they only yield values when explicitly asked.
  • You can feed the output of a generator to the input of another generator to form data pipelines.
  • Data pipelines can be modularized and customized to your needs.
  • Generators are useful for generating values ad infinitum.


Generators don't have to be complicated topics, they have a place in any Python programmer's repertoire if given the time to be understood. Even in Big Data situations where simpler methods fall short, generator-based analyses still stand tall. There's a lot about generators that we didn't discuss here, but it should still give you a good foundation to start using them in your own analytic life.

Further Resources:

Christian Pascual

About the author

Christian Pascual

Christian is a PhD student studying biostatistics in California. He enjoys making statistics and programming more accessible to a wider audience. Outside of school, he enjoys going to the gym, language learning, and woodworking.