Functional programming

Functional programming#

Mutability and state#

We talked a bit about mutability in Python already. Python’s concept of mutability has two aspects:

Some built-in (implemented in C) objects are immutable. You can’t really imitate this in only Python with custom classes - all user classes make mutable objects.
Immutable objects have a hash (__hash__ is not None). This is the actual property checked most of the time - so you can make classes that can be put in sets and dict keys by making them hashable. It is up to you to make them immutable by convention (you can make it pretty hard to do).

Notice the second point: immutability really is by convention, not forced by the language - and that’s okay. Immutability / mutability is a design pattern!

Why do we tend toward mutable?#

Mutable is easy to change a small part#

It’s conceptually and syntactically easy to change a small part of a structure. For example:

import dataclasses

@dataclasses.dataclass
class Mutable:
    x: int
    y: int

mutable = Mutable(1, 2)
mutable.x = 2
print(mutable)

Mutable(x=2, y=2)

If we wanted to change this without mutation, we’d have to do:

@dataclasses.dataclass(frozen=True)
class Immutable:
    x: int
    y: int

immutable_1 = Immutable(1, 2)
immutable_2 = Immutable(2, immutable_1.y)
print(immutable_2)

Immutable(x=2, y=2)

Imagine if the class was larger! Dataclasses can really help us here in the immutable case write something that is quite reasonable instead of having to manually call the constructor ourselves with the replaced values and forwarding in all the rest:

immutable_1 = Immutable(1, 2)
immutable_2 = dataclasses.replace(immutable_1, x=2)
print(immutable_2)

Immutable(x=2, y=2)

This scales much better to larger dataclasses. But it’s still a bit easier / more natural, and much easier if you are not dealing with something like dataclasses (though as we’ll see, things that tend heavily toward immutability will provide tools similar to dataclasses.replace)

Memory saving#

If you have a large structure, you can simply change part in place, avoiding copying the entire thing. This is nice - but it could be seen as an implementation detail if you never were to touch the original structure again. If the language was smart enough to detect this, it could do the mutation for you even when you asked for a new copy. That is, the difference the examples above is really an implementation detail if immutable_1 is never used again. A smart compiler (not Python, which doesn’t have a compiler) could learn to rewrite the immutable example into the first behind the scenes.

We’ll get into why we think the immutable versions may be better in some cases soon!

Easy to build API in a way that you maybe shouldn’t#

Some APIs just don’t fit well with immutable designs (we’ll see some solutions, though). Things with a non-linear progression, for example. Something like:

data = Data()
data.load_data()
data.prepare()
data.do_calculations()
data.plot()

That’s easy and simple. Unless you forget a step. Oh, yeah, and static analysis tools can’t tell you if you forget a step, the API doesn’t statically “know” that .prepare() is required, for example. Tab completion tells you that .plot() is valid immediately. The problem is that Data has a changing state, and not all operations are valid in all possible states.

If we replaced the single Data with multiple immutable classes, then our problem would be solved:

empty_data = EmptyData()
loaded_data = empty_data.load_data()
prepared_data = loaded_data.prepare()
computed_data = prepared_data.do_calculations()
computed_data.plot()

These classes don’t have to be immutable. Maybe you can load more data to loaded data. But they are harder to use incorrectly when they at least not have a mutating state that makes subsets of operations invalid. Note that tab completion in this case would show exactly the allowed set of operations each time.

We could avoid naming the temporaries, too:

computed_data = EmptyData().load_data().prepare().do_calculations()
computed_data.plot()

This “chaining” style is a hallmark of functional programming (which is where we are headed).

Copy vs. reference#

As you might already know, everything in Python is copied by reference. If you have a mutable object, you have to make a copy (or a deepcopy) to ensure that its contents (or contents of its contents) are not changed. For example, this function is evil:

def evil(x):
    x.append("Muhahaha")

mutable = []
evil(mutable)
print(mutable)

['Muhahaha']

A function should be of the form name(input, ...) -> output, but instead, this function is mutating the argument it was given! If we used a immutable object instead, we would have been forced to use the return value of the function, making it much easier to understand when you are using it. Functions are much nicer when arguments are not mutated. self is kind of a special case - a programmer is much more likely to expect x.do_something() to modify x than do_something(x). But as we’ve seen above, there are limits/drawbacks.

What do we get with immutability?#

Let’s summarize above:

Optimization choices for the compiler (if you have one) - it’s much simpler to reason about.
Chaining of methods
No worry about copying

Notice I didn’t actually require immutability on the points, I simply required we don’t mutate. The fact that we can or can’t mutate it is less important than if we do mutate it.

Functional programming#

Let’s define a pure function. This is a function that:

Does not mutate its arguments
Does not contain internal state (doesn’t mutate itself or a global, basically)
Has no side effects (like printing to the screen)

Functors are not pure functions (they mutate themselves). print is not a pure function (it mutates the screen). And many of the methods on lists and dicts are not pure functions (they mutate the first argument, self). But there are lots of pure functions, like most built-ins and most non-method functions, and methods of tuples and such.

Put simply, a pure function produces the same output whenever given the same inputs. Its behavior is determined solely by the inputs.

Map, filter, reduce#

Functional programming often involves passing functions to functions. Three very common ones are map (apply a function to each item of a sequence), filter (remove items from a sequence based on a function), and reduce (apply a function to successive pairs of a sequence). You’ll sometimes see map called apply, and reduce called fold.

Let’s take the following Pythonic code using a comprehension:

items = [1, 2, 3, 4, 5]
sum_sq_odds = sum(x**2 for x in items if x % 2 == 1)
print(sum_sq_odds)

And instead write it following in a functional style:

import functools


items = [1, 2, 3, 4, 5]
sum_sq_odds = functools.reduce(lambda x, y: x + y, filter(lambda x: x % 2 == 1, map(lambda x: x**2, items)))
print(sum_sq_odds)

Notice the data structure I chose was a list, which is mutable - but that’s okay, I didn’t mutate it - the functions were pure.

Notice how horribly this reads; our operations are in reverse order (right to left). In a more functional language, you can chain these left to right instead. Or with a library.

We’ll write a little wrapper that will allow us to chain operations as if we were in a proper functional language:

class FunctionalIterable:
    def __init__(self, this, /):
        self._this = this
    def __repr__(self):
        return repr(self._this)
    def map(self, func):
        return self.__class__(map(func, self._this))
    def filter(self, func):
        return self.__class__(filter(func, self._this))
    def reduce(self, func):
        return functools.reduce(func, self._this)

Now, let’s compare readability:

items = FunctionalIterable([1, 2, 3, 4, 5])
sum_sq_odds = items.map(lambda x: x**2).filter(lambda x: x % 2).reduce(lambda x,y: x + y)
print(sum_sq_odds)

Notice that we now read this left to right. Our items are squared then filtered then reduced. Compare that to some languages with support for functional programming:

Ruby

items = [1,2,3,4,5]
sum_sq_odds = items.map{_1**2}.filter{_1 % 2 == 1}.reduce{_1 + _2}
puts items

This looks very nice, especially with Ruby 2.7’s lambda shortcut notation. If you didn’t want to use that, classic lambdas would be |x| x**2, etc.

Scala

val arr = (1 to 5)
val sum_sq_odds = arr.map(x => x*x).filter(_ % 2 == 1).reduce(_ + _)
println(sum_sq_odds)

This is often considered a highly functional language, so I’ve included it here for completeness.

Hascal

foldl (+) 0 . filter ((==) 1 . flip mod 2) . map (^2) $ [1..5]

Hascal is a purely functional language. It’s also a bit odd in that idiomatic Hascal reads right to left (from mathematics). However, as of 4.8 it does have reverse function application operator (also using a lambda to match the other examples more closely):

[1..5] & map (^2) & filter (\x -> x `mod` 2 == 1) & foldl (+) 0

I’ve left of variable assignment & printing since those are a bit different in online playgrounds.

C++20 Ranges

#include <iostream>
#include <vector>
#include <ranges>
#include <numeric>

int main() {
    std::vector<int> items {1, 2, 3, 4, 5};
    auto odd_sq = items | std::views::transform([](int i){return i*i;})
                        | std::views::filter([](int i){return i%2==1;});
    sum_sq_odds = std::accumulate(std::begin(odd_sq), std::end(odd_sq), 0, [](int a, int b){return a + b;})
    std::cout << result << std::endl;
    return 0;
}

Not that C++ is slowly gaining support; std::fold_left is in C++23, but for C++20, we have to drop back to a classic std::accumulate algorithm & begin and end iterators.

Rust

fn main() {
    let items = [1,2,3,4,5];
    let sum_sq_odds = items.iter()
                      .map(|x| x*x)
                      .filter(|&x| x%2==1)
                      .fold(0, |acc, x| acc + x);
    println!("{}", sum_sq_odds);
}

Notice how many of the languages use similar terms and try to read well when chained. Map (or transform for C++, odd one out here) applies a function to an iterator and returns an iterator of the returned values. Filter removes iterations based on the truthiness of the return value. And reduce (or fold) does a binary reduction (most languages also have sum for this common reduction).

Most templating languages like Liquid and Jinja have “filters”, which are applied left to right in a similar style. Hugo’s templating system is functional, too.

Currying#

Another common feature of functional languages is currying. Let’s say a function takes two arguments. Currying means that if you give it one argument, you now have a function that requires one argument. Purely functional or heavily functional languages will often curry by default; in Python, you have to manually curry with functools.partial.

Here’s an example of a two argument function:

def power(y, x):
    return x**y

Now, we can “curry it”:

import functools


pow2 = functools.partial(power, 2)
pow3 = functools.partial(power, 3)

Now we have two new functions that have y pre-specified:

print(f"{pow2(10) = }")
print(f"{pow3(10) = }")

pow2(10) = 100

pow3(10) = 1000

Note that this is a (better) way to do a subset of Functors (from the previous chapter on OOP). If you are using a Functor to capture state, then you can use currying with partial to be much more explicit as to your intent and curried methods will capture the current value, not a reference that could change later.

So what do you get?#

Why worry about pure functions? It turns out, they have a lot of properties related to optimization. Some libraries make use of these properties to do some very impressive things.

Lazy#

When you have pure functions, the programming language or library can decide when to run them - and often it can decide to do it at some later time. For example, we can observe this if we break our promise about pure functions:

def not_pure(x):
    print("computing x")
    return x

values = [1,2,3]
results = map(not_pure, values)

Notice nothing ran? If you look at the type of results, it’s an iterator. The mapping only happens when you do the iteration. We can also filter:

filtered_results = filter(lambda x: x%2==0, results)

Still nothing. A reduction does iterate over the results:

print(sum(filtered_results))

computing x

computing x

computing x

Now you see the side effects; this is when it ran.

Python & lazy iterators

You can do this without the functional spellings above - generator comprehensions (the ones with () instead of []) are also lazy, as is range and quite a few other methods in Python 3. There also is good support for coroutines via generator functions. You can use those to write things that look like they apply to a list and return a list but still ensure code only runs once though a final iteration, saving memory. However, this is not useful for most numeric work - see array programming.

Parallelization#

While not used in Python, functional programming is easy to optimize for parallel execution. Since things like external state are not allowed, you can not just run later, you can also run concurrently.

Easier to optimize#

Some languages and libraries can do a lot of optimization if you have pure functions and no mutability. There are a lot of assumptions you can make that help with optimization. Again, not in Python itself, but we will look at a library that takes advantage of these things.

Example: JAX#

No JAX in our environment

We will not be including or running JAX since it is quite large and would slow down setting up an environment. All examples will be pre-computed, and we’ll not be using it in homework.

A library that makes use of this is JAX. JAX is a ML inspired library that is a bit of a spiritual successor to TensorFlow. TensorFlow focused on symbolic computation, while JAX is functional.

JAX provides a NumPy-like interface:

import jax
import jax.numpy as jnp

arr = jnp.array([1, 2, 3])

However, you can’t modify a jax array. Having read the above, I would expect you could have guessed that JAX is functional. They provide a shortcut for setting a value; for example, arr[0] = 4 can be written:

arr = arr.at[0].set(4)

But what do you get?

You can fuse functions together with jax.jit & they get compiled into machine code
You can target CPU, GPU, and TPU (Google’s tensor processing units)
You can compute gradients of functions

For example, we can define a pure function and get a fast version and get a gradient function as well:

@jax.jit
def f(x):
    return x**3 + x**2 + x


dfdx = jax.grad(f)

And we can use it:

dfdx(1.0)

DeviceArray(6., dtype=float32, weak_type=True)

(That’s \(\frac{df}{dx} = 3x^2 + 2x + 1\) if you were to compute it by hand.)

There are quite a few rules to follow when making JAX functions (control flow is a common issue), but you get a lot for your efforts!

Other ways to go faster

You don’t have to use JAX to write code faster than NumPy. There are other tools that sometimes are better in some situations, including the excellent Numba library, which is imperative. PyTorch is a great ML-focused library. CuPy is a great GPU NumPy replacement. Etc. JAX here is just intended to be an example of what thinking in a functional mindset can do.

Functional programming

Contents

Functional programming#

Mutability and state#

Why do we tend toward mutable?#

Mutable is easy to change a small part#

Memory saving#

Easy to build API in a way that you maybe shouldn’t#

Copy vs. reference#

What do we get with immutability?#

Functional programming#

Map, filter, reduce#

Currying#

So what do you get?#

Lazy#

Parallelization#

Easier to optimize#

Example: JAX#