Cool Python features for machine learning that I used to be too afraid to use. Will be updated as I have more time / learn more.

python-is-cool

A gentle guide to the Python features that I didn’t know existed or was too afraid to use. This will be updated as I learn more and become less lazy.

This uses python >= 3.6.

GitHub has problem rendering Jupyter notebook so I copied the content here. I still keep the notebook in case you want to clone and run it on your machine, but you can also click the Binder badge below and run it in your browser.

1. Lambda, map, filter, reduce

The lambda keyword is used to create inline functions. The functionssquare_fn and square_ld below are identical.

def square_fn(x):
    return x * x

square_ld = lambda x: x * x

for i in range(10):
    assert square_fn(i) == square_ld(i)

Its quick declaration makes lambda functions ideal for use in callbacks, and when functions are to be passed as arguments to other functions. They are especially useful when used in conjunction with functions like map, filter, and reduce.

map(fn, iterable) applies the fn to all elements of the iterable (e.g. list, set, dictionary, tuple, string) and returns a map object.

nums = [1/3, 333/7, 2323/2230, 40/34, 2/3]
nums_squared = [num * num for num in nums]
print(nums_squared)

==> [0.1111111, 2263.04081632, 1.085147, 1.384083, 0.44444444]

This is the same as calling using map with a callback function.

nums_squared_1 = map(square_fn, nums)
nums_squared_2 = map(lambda x: x * x, nums)
print(list(nums_squared_1))

==> [0.1111111, 2263.04081632, 1.085147, 1.384083, 0.44444444]

You can also use map with more than one iterable. For example, if you want to calculate the mean squared error of a simple linear function f(x) = ax + b with the true label labels, these two methods are equivalent:

a, b = 3, -0.5
xs = [2, 3, 4, 5]
labels = [6.4, 8.9, 10.9, 15.3]

# Method 1: using a loop
errors = []
for i, x in enumerate(xs):
    errors.append((a * x + b - labels[i]) ** 2)
result1 = sum(errors) ** 0.5 / len(xs)

# Method 2: using map
diffs = map(lambda x, y: (a * x + b - y) ** 2, xs, labels)
result2 = sum(diffs) ** 0.5 / len(xs)

print(result1, result2)

==> 0.35089172119045514 0.35089172119045514

Note that objects returned by map and filter are iterators, which means that their values aren’t stored but generated as needed. After you’ve called sum(diffs), diffs becomes empty. If you want to keep all elements in diffs, convert it to a list using list(diffs).

filter(fn, iterable) works the same way as map, except that fn returns a boolean value and filter returns all the elements of the iterable for which the fn returns True.

bad_preds = filter(lambda x: x > 0.5, errors)
print(list(bad_preds))

==> [0.8100000000000006, 0.6400000000000011]

reduce(fn, iterable, initializer) is used when we want to iteratively apply an operator to all elements in a list. For example, if we want to calculate the product of all elements in a list:

product = 1
for num in nums:
    product *= num
print(product)

==> 12.95564683272412

This is equivalent to:

from functools import reduce
product = reduce(lambda x, y: x * y, nums)
print(product)

==> 12.95564683272412

Note on the performance of lambda functions

Lambda functions are meant for one time use. Each time lambda x: dosomething(x) is called, the function has to be created, which hurts the performance if you call lambda x: dosomething(x) multiple times (e.g. when you pass it inside reduce).

When you assign a name to the lambda function as in fn = lambda x: dosomething(x), its performance is slightly slower than the same function defined using def, but the difference is negligible. See here.

Even though I find lambdas cool, I personally recommend using named functions when you can for the sake of clarity.

2. List manipulation

Python lists are super cool.

2.1 Unpacking

We can unpack a list by each element like this:

elems = [1, 2, 3, 4]
a, b, c, d = elems
print(a, b, c, d)

==> 1 2 3 4

We can also unpack a list like this:

a, *new_elems, d = elems
print(a)
print(new_elems)
print(d)

==> 1
    [2, 3]
    4

2.2 Slicing

We know that we can reverse a list using [::-1].

elems = list(range(10))
print(elems)

==> [0, 1, 2, 3, 4, 5, 6, 7, 8, 9]

print(elems[::-1])

==> [9, 8, 7, 6, 5, 4, 3, 2, 1, 0]

The syntax [x:y:z] means "take every zth element of a list from index x to index y". When z is negative, it indicates going backwards. When x isn’t specified, it defaults to the first element of the list in the direction you are traversing the list. When y isn’t specified, it defaults to the last element of the list. So if we want to take every 2th element of a list, we use [::2].

evens = elems[::2]
print(evens)

reversed_evens = elems[-2::-2]
print(reversed_evens)

==> [0, 2, 4, 6, 8]
    [8, 6, 4, 2, 0]

We can also use slicing to delete all the even numbers in the list.

del elems[::2]
print(elems)

==> [1, 3, 5, 7, 9]

2.3 Insertion

We can change the value of an element in a list to another value.

elems = list(range(10))
elems[1] = 10
print(elems)

==> [0, 10, 2, 3, 4, 5, 6, 7, 8, 9]

If we want to replace the element at an index with multiple elements, e.g. replace the value 1 with 3 values 20, 30, 40:

elems = list(range(10))
elems[1:2] = [20, 30, 40]
print(elems)

==> [0, 20, 30, 40, 2, 3, 4, 5, 6, 7, 8, 9]

If we want to insert 3 values 0.2, 0.3, 0.5 between element at index 0 and element at index 1:

elems = list(range(10))
elems[1:1] = [0.2, 0.3, 0.5]
print(elems)

==> [0, 0.2, 0.3, 0.5, 1, 2, 3, 4, 5, 6, 7, 8, 9]

2.4 Flattening

We can flatten a list of lists using sum.

list_of_lists = [[1], [2, 3], [4, 5, 6]]
sum(list_of_lists, [])

==> [1, 2, 3, 4, 5, 6]

If we have nested lists, we can recursively flatten it. That’s another beauty of lambda functions – we can use it in the same line as its creation.

nested_lists = [[1, 2], [[3, 4], [5, 6], [[7, 8], [9, 10], [[11, [12, 13]]]]]]
flatten = lambda x: [y for l in x for y in flatten(l)] if type(x) is list else [x]
flatten(nested_lists)

# This line of code is from
# https://github.com/sahands/python-by-example/blob/master/python-by-example.rst#flattening-lists

2.5 List vs generator

To illustrate the difference between a list and a generator, let’s look at an example of creating n-grams out of a list of tokens.

One way to create n-grams is to use a sliding window.

tokens = ['i', 'want', 'to', 'go', 'to', 'school']

def ngrams(tokens, n):
    length = len(tokens)
    grams = []
    for i in range(length - n + 1):
        grams.append(tokens[i:i+n])
    return grams

print(ngrams(tokens, 3))

==> [['i', 'want', 'to'],
     ['want', 'to', 'go'],
     ['to', 'go', 'to'],
     ['go', 'to', 'school']]

In the above example, we have to store all the n-grams at the same time. If the text has m tokens, then the memory requirement is O(nm), which can be problematic when m is large.

Instead of using a list to store all n-grams, we can use a generator that generates the next n-gram when it’s asked for. This is known as lazy evaluation. We can make the function ngrams returns a generator using the keyword yield. Then the memory requirement is O(m+n).

def ngrams(tokens, n):
    length = len(tokens)
    for i in range(length - n + 1):
        yield tokens[i:i+n]

ngrams_generator = ngrams(tokens, 3)
print(ngrams_generator)

==> <generator object ngrams at 0x1069b26d0>

for ngram in ngrams_generator:
    print(ngram)

==> ['i', 'want', 'to']
    ['want', 'to', 'go']
    ['to', 'go', 'to']
    ['go', 'to', 'school']

Another way to generate n-grams is to use slices to create lists: [0, 1, ..., -n], [1, 2, ..., -n+1], …, [n-1, n, ..., -1], and then zip them together.

def ngrams(tokens, n):
    length = len(tokens)
    slices = (tokens[i:length-n+i+1] for i in range(n))
    return zip(*slices)

ngrams_generator = ngrams(tokens, 3)
print(ngrams_generator)

==> <zip object at 0x1069a7dc8> # zip objects are generators

for ngram in ngrams_generator:
    print(ngram)

==> ('i', 'want', 'to')
    ('want', 'to', 'go')
    ('to', 'go', 'to')
    ('go', 'to', 'school')

Note that to create slices, we use (tokens[...] for i in range(n)) instead of [tokens[...] for i in range(n)]. [] is the normal list comprehension that returns a list. () returns a generator.

3. Classes and magic methods

In Python, magic methods are prefixed and suffixed with the double underscore __, also known as dunder. The most wellknown magic method is probably __init__.

class Node:
    """ A struct to denote the node of a binary tree.
    It contains a value and pointers to left and right children.
    """
    def __init__(self, value, left=None, right=None):
        self.value = value
        self.left = left
        self.right = right

When we try to print out a Node object, however, it’s not very interpretable.

root = Node(5)
print(root) # <__main__.Node object at 0x1069c4518>

Ideally, when user prints out a node, we want to print out the node’s value and the values of its children if it has children. To do so, we use the magic method __repr__, which must return a printable object, like string.

class Node:
    """ A struct to denote the node of a binary tree.
    It contains a value and pointers to left and right children.
    """
    def __init__(self, value, left=None, right=None):
        self.value = value
        self.left = left
        self.right = right

    def __repr__(self):
        strings = [f'value: {self.value}']
        strings.append(f'left: {self.left.value}' if self.left else 'left: None')
        strings.append(f'right: {self.right.value}' if self.right else 'right: None')
        return ', '.join(strings)

left = Node(4)
root = Node(5, left)
print(root) # value: 5, left: 4, right: None

We’d also like to compare two nodes by comparing their values. To do so, we overload the operator == with __eq__, < with __lt__, and >= with __ge__.

class Node:
    """ A struct to denote the node of a binary tree.
    It contains a value and pointers to left and right children.
    """
    def __init__(self, value, left=None, right=None):
        self.value = value
        self.left = left
        self.right = right

    def __eq__(self, other):
        return self.value == other.value

    def __lt__(self, other):
        return self.value < other.value

    def __ge__(self, other):
        return self.value >= other.value


left = Node(4)
root = Node(5, left)
print(left == root) # False
print(left < root) # True
print(left >= root) # False

For a comprehensive list of supported magic methods here or see the official Python documentation here (slightly harder to read).

Some of the methods that I highly recommend:

• __len__: to overload the len() function.
• __str__: to overload the str() function.
• __iter__: if you want to your objects to be iterators. This also allows you to call next() on your object.

For classes like Node where we know for sure all the attributes they can support (in the case of Node, they are value, left, and right), we might want to use __slots__ to denote those values for both performance boost and memory saving. For a comprehensive understanding of pros and cons of __slots__, see this absolutely amazing answer by Aaron Hall on StackOverflow.

class Node:
    """ A struct to denote the node of a binary tree.
    It contains a value and pointers to left and right children.
    """
    __slots__ = ('value', 'left', 'right')
    def __init__(self, value, left=None, right=None):
        self.value = value
        self.left = left
        self.right = right

4. local namespace, object’s attributes

The locals() function returns a dictionary containing the variables defined in the local namespace.

class Model1:
    def __init__(self, hidden_size=100, num_layers=3, learning_rate=3e-4):
        print(locals())
        self.hidden_size = hidden_size
        self.num_layers = num_layers
        self.learning_rate = learning_rate

model1 = Model1()

==> {'learning_rate': 0.0003, 'num_layers': 3, 'hidden_size': 100, 'self': <__main__.Model1 object at 0x1069b1470>}

All attributes of an object are stored in its __dict__.

print(model1.__dict__)

==> {'hidden_size': 100, 'num_layers': 3, 'learning_rate': 0.0003}

Note that manually assigning each of the arguments to an attribute can be quite tiring when the list of the arguments is large. To avoid this, we can directly assign the list of arguments to the object’s __dict__.

class Model2:
    def __init__(self, hidden_size=100, num_layers=3, learning_rate=3e-4):
        params = locals()
        del params['self']
        self.__dict__ = params

model2 = Model2()
print(model2.__dict__)

==> {'learning_rate': 0.0003, 'num_layers': 3, 'hidden_size': 100}

This can be especially convenient when the object is initiated using the catch-all **kwargs, though the use of **kwargs should be reduced to the minimum.

class Model3:
    def __init__(self, **kwargs):
        self.__dict__ = kwargs

model3 = Model3(hidden_size=100, num_layers=3, learning_rate=3e-4)
print(model3.__dict__)

==> {'hidden_size': 100, 'num_layers': 3, 'learning_rate': 0.0003}

5. Wild import

Often, you run into this wild import * that looks something like this:

file.py

    from parts import *

This is irresponsible because it will import everything in module, even the imports of that module. For example, if parts.py looks like this:

parts.py

import numpy
import tensorflow

class Encoder:
    ...

class Decoder:
    ...

class Loss:
    ...

def helper(*args, **kwargs):
    ...

def utils(*args, **kwargs):
    ...

Since parts.py doesn’t have __all__ specified, file.py will import Encoder, Decoder, Loss, utils, helper together with numpy and tensorflow.

If we intend that only Encoder, Decoder, and Loss are ever to be imported and used in another module, we should specify that in parts.py using the __all__ keyword.

parts.py

 __all__ = ['Encoder', 'Decoder', 'Loss']
import numpy
import tensorflow

class Encoder:
    ...

Now, if some user irresponsibly does a wild import with parts, they can only import Encoder, Decoder, Loss. Personally, I also find __all__ helpful as it gives me an overview of the module.

6. Decorator to time your functions

It’s often useful to know how long it takes a function to run, e.g. when you need to compare the performance of two algorithms that do the same thing. One naive way is to call time.time() at the begin and end of each function and print out the difference.

For example: compare two algorithms to calculate the n-th Fibonacci number, one uses memoization and one doesn’t.

def fib_helper(n):
    if n < 2:
        return n
    return fib_helper(n - 1) + fib_helper(n - 2)

def fib(n):
    """ fib is a wrapper function so that later we can change its behavior
    at the top level without affecting the behavior at every recursion step.
    """
    return fib_helper(n)

def fib_m_helper(n, computed):
    if n in computed:
        return computed[n]
    computed[n] = fib_m_helper(n - 1, computed) + fib_m_helper(n - 2, computed)
    return computed[n]

def fib_m(n):
    return fib_m_helper(n, {0: 0, 1: 1})

Let’s make sure that fib and fib_m are functionally equivalent.

for n in range(20):
    assert fib(n) == fib_m(n)

import time

start = time.time()
fib(30)
print(f'Without memoization, it takes {time.time() - start:7f} seconds.')

==> Without memoization, it takes 0.267569 seconds.

start = time.time()
fib_m(30)
print(f'With memoization, it takes {time.time() - start:.7f} seconds.')

==> With memoization, it takes 0.0000713 seconds.

If you want to time multiple functions, it can be a drag having to write the same code over and over again. It’d be nice to have a way to specify how to change any function in the same way. In this case would be to call time.time() at the beginning and the end of each function, and print out the time difference.

This is exactly what decorators do. They allow programmers to change the behavior of a function or class. Here’s an example to create a decorator timeit.

def timeit(fn): 
    # *args and **kwargs are to support positional and named arguments of fn
    def get_time(*args, **kwargs): 
        start = time.time() 
        output = fn(*args, **kwargs)
        print(f"Time taken in {fn.__name__}: {time.time() - start:.7f}")
        return output  # make sure that the decorator returns the output of fn
    return get_time 

Add the decorator @timeit to your functions.

@timeit
def fib(n):
    return fib_helper(n)

@timeit
def fib_m(n):
    return fib_m_helper(n, {0: 0, 1: 1})

fib(30)
fib_m(30)

==> Time taken in fib: 0.2787242
==> Time taken in fib_m: 0.0000138

7. Caching with @functools.lru_cache

Memoization is a form of cache: we cache the previously calculated Fibonacci numbers so that we don’t have to calculate them again.

Caching is such an important technique that Python provides a built-in decorator to give your function the caching capacity. If you want fib_helper to reuse the previously calculated Fibonacci numbers, you can just add the decorator lru_cache from functools. lru stands for “least recently used”. For more information on cache, see here.

import functools

@functools.lru_cache()
def fib_helper(n):
    if n < 2:
        return n
    return fib_helper(n - 1) + fib_helper(n - 2)

@timeit
def fib(n):
    """ fib is a wrapper function so that later we can change its behavior
    at the top level without affecting the behavior at every recursion step.
    """
    return fib_helper(n)

fib(50)
fib_m(50)

==> Time taken in fib: 0.0000412
==> Time taken in fib_m: 0.0000281