Migrating to Python 3 with pleasure

A short guide on features of Python 3 for data scientists

Python became a mainstream language for machine learning and other scientific fields that heavily operate with data; it boasts various deep learning frameworks and well-established set of tools for data processing and visualization.

However, Python ecosystem co-exists in Python 2 and Python 3, and Python 2 is still used among data scientists. By the end of 2019 scientific stack will stop supporting Python2. As for numpy, after 2018 any new feature releases will support only Python3.

To make transition less frustrating, I've collected a bunch of Python 3 features that you may find useful.

Image from Dario Bertini post (toptal)

Better paths handling with pathlib

pathlib is a default module in python3, that helps you to avoid tons of os.path.joins:

from pathlib import Path

dataset = 'wiki_images'
datasets_root = Path('/path/to/datasets/') 

train_path = datasets_root / dataset / 'train'
test_path = datasets_root / dataset / 'test'

for image_path in train_path.iterdir():
    with image_path.open() as f: # note, open is a method of Path object
        # do something with an image

Previously it was always tempting to use string concatenation (concise, but obviously bad), now with pathlib the code is safe, concise, and readable.

Also pathlib.Path has a bunch of methods, that every python novice previously had to google:

p.exists()
p.is_dir()
p.parts()
p.with_name('sibling.png') # only change the name, but keep the folder
p.with_suffix('.jpg') # only change the extension, but keep the folder and the name
p.chmod(mode)
p.rmdir()

pathlib should save you lots of time, please see docs and reference for more.

Type hinting is now part of the language

Example of type hinting in pycharm:

Python is not just a language for small scripts anymore, data pipelines these days include numerous steps each involving different frameworks (and sometimes very different logic).

Type hinting was introduced to help with growing complexity of programs, so machines could help with code verification. Previously different modules used custom ways to point types in doctrings (Hint: pycharm can convert old docstrings to fresh typehinting).

As a simple example, the following code may work with different types of data (that's what we like about python data stack).

def repeat_each_entry(data):
    """ Each entry in the data is doubled 
    <blah blah nobody reads the documentation till the end>
    """
    index = numpy.repeat(numpy.arange(len(data)), 2)
    return data[index]

This code e.g. works for numpy.array (incl. multidimensional ones), astropy.Table and astropy.Column, bcolz, cupy, mxnet.ndarray and others.

This code will work for pandas.Series, but in the wrong way:

repeat_each_entry(pandas.Series(data=[0, 1, 2], index=[3, 4, 5])) # returns Series with Nones inside

This was code of two lines. Imagine how unpredictable behavior of a complex system, because just one function may misbehave. Putting explicitly which types method expects is very helpful in large systems, this will warn you if function wasn't expected to get such arguments.

def repeat_each_entry(data: Union[numpy.ndarray, bcolz.carray]):

If you have a significant codebase, hinting tools like MyPy are likely to become part of your continuous integration pipeline. A webinar "Putting Type Hints to Work" by Daniel Pyrathon is good for a brief introduction.

Sidenote: unfortunately, hinting is not yet powerful enough to provide fine-grained typing for ndarrays/tensors, but maybe we'll have it once, and this will be a great feature for DS.

Type hinting → type checking in runtime

By default, function annotations do not influence how your code is working, but merely help you to point code intentions.

However, you can enforce type checking in runtime with tools like ... enforce, this can help you in debugging (there are many cases when type hinting is not working).

@enforce.runtime_validation
def foo(text: str) -> None:
    print(text)

foo('Hi') # ok
foo(5)    # fails


@enforce.runtime_validation
def any2(x: List[bool]) -> bool:
    return any(x)

any ([False, False, True, False]) # True
any2([False, False, True, False]) # True

any (['False']) # True
any2(['False']) # fails

any ([False, None, "", 0]) # False
any2([False, None, "", 0]) # fails

Other usages of function annotations

As mentioned before, annotations do not influence code execution, but rather provide some meta-information, and you can use it as you wish.

For instance, measure units is a common pain in scientific areas, astropy package provides a simple decorator to control units of input quantities and convert output to required units

# Python 3
from astropy import units as u
@u.quantity_input()
def frequency(speed: u.meter / u.s, wavelength: u.m) -> u.terahertz:
    return speed / wavelength
    
frequency(speed=300_000 * u.km / u.s, wavelength=555 * u.nm)
# output: 540.5405405405404 THz, frequency of green visible light

If you're processing tabular scientific data in python (not necessarily astronomical), you should give astropy a shot.

You can also define your application-specific decorators to perform control / conversion of inputs and output in the same manner.

Matrix multiplication with @

Let's implement one of the simplest ML models — a linear regression with l2 regularization (a.k.a. ridge regression):

# l2-regularized linear regression: || AX - b ||^2 + alpha * ||x||^2 -> min

# Python 2
X = np.linalg.inv(np.dot(A.T, A) + alpha * np.eye(A.shape[1])).dot(A.T.dot(b))
# Python 3
X = np.linalg.inv(A.T @ A + alpha * np.eye(A.shape[1])) @ (A.T @ b)

The code with @ becomes more readable and more translatable between deep learning frameworks: same code X @ W + b[None, :] for a single layer of perceptron works in numpy, cupy, pytorch, tensorflow (and other frameworks that operate with tensors).

Globbing with **

Recursive folder globbing is not easy in Python 2, even custom module glob2 exists that overcomes this. Recursive flag is supported since Python 3.6:

import glob

# Python 2
found_images = \
    glob.glob('/path/*.jpg') \
  + glob.glob('/path/*/*.jpg') \
  + glob.glob('/path/*/*/*.jpg') \
  + glob.glob('/path/*/*/*/*.jpg') \
  + glob.glob('/path/*/*/*/*/*.jpg') 

# Python 3
found_images = glob.glob('/path/**/*.jpg', recursive=True)

Better option is to use pathlib in python3 (minus one import!):

# Python 3
found_images = pathlib.Path('/path/').glob('**/*.jpg')

Print is a function now

Yes, code now has these annoying parentheses, but there are some advantages:

  • simple syntax for using file descriptor:

    print >>sys.stderr, "critical error"      # Python 2
    print("critical error", file=sys.stderr)  # Python 3
  • printing tab-aligned tables without str.join:

    # Python 3
    print(*array, sep='\t')
    print(batch, epoch, loss, accuracy, time, sep='\t')
  • hacky suppressing / redirection of printing output:

    # Python 3
    _print = print # store the original print function
    def print(*args, **kargs):
        pass  # do something useful, e.g. store output to some file

    In jupyter it is desirable to log each output to a separate file (to track what's happening after you got disconnected), so you can override print now.

    Below you see a context manager that temporarily overrides behavior of print:

    @contextlib.contextmanager
    def replace_print():
        import builtins
        _print = print # saving old print function
        # or use some other function here
        builtins.print = lambda *args, **kwargs: _print('new printing', *args, **kwargs)
        yield
        builtins.print = _print
    
    with replace_print():
        <code here will invoke other print function>

    It is not a recommended approach, but a small dirty hack that is now possible.

  • print can participate in list comprehensions and other language constructs

    # Python 3
    result = process(x) if is_valid(x) else print('invalid item: ', x)

f-strings for simple and reliable formatting

Default formatting system provides a flexibility that is not required in data experiments. Resulting code is either too verbose or too fragile towards any changes.

Quite typically data scientist outputs iteratively some logging information in a fixed format. It is common to have a code like:

# Python 2
print('{batch:3} {epoch:3} / {total_epochs:3}  accuracy: {acc_mean:0.4f}±{acc_std:0.4f} time: {avg_time:3.2f}'.format(
    batch=batch, epoch=epoch, total_epochs=total_epochs, 
    acc_mean=numpy.mean(accuracies), acc_std=numpy.std(accuracies),
    avg_time=time / len(data_batch)
))

# Python 2 (too error-prone during fast modifications, please avoid):
print('{:3} {:3} / {:3}  accuracy: {:0.4f}±{:0.4f} time: {:3.2f}'.format(
    batch, epoch, total_epochs, numpy.mean(accuracies), numpy.std(accuracies),
    time / len(data_batch)
))

Sample output:

120  12 / 300  accuracy: 0.8180±0.4649 time: 56.60

f-strings aka formatted string literals were introduced in Python 3.6:

# Python 3.6+
print(f'{batch:3} {epoch:3} / {total_epochs:3}  accuracy: {numpy.mean(accuracies):0.4f}±{numpy.std(accuracies):0.4f} time: {time / len(data_batch):3.2f}')

Also, very handy when writing queries or generate code fragments:

query = f"INSERT INTO STATION VALUES (13, '{city}', '{state}', {latitude}, {longitude})"

Important: don't forget to escape arguments to prevent SQL injection attacks.

Explicit difference between 'true division' and 'integer division'

For data science this is definitely a handy change (but not for system programming, I believe)

data = pandas.read_csv('timing.csv')
velocity = data['distance'] / data['time']

Result in Python 2 depends on whether 'time' and 'distance' (e.g. measured in meters and seconds) are stored as integers. In Python 3, result is correct in both cases, because result of division is float.

Another case is integer division, which is now an explicit operation:

n_gifts = money // gift_price  # correct for int and float arguments

Note, that this applies both to built-in types and to custom types provided by data packages (e.g. numpy or pandas).

Strict ordering

# All these comparisons are illegal in Python 3
3 < '3'
2 < None
(3, 4) < (3, None)
(4, 5) < [4, 5]

# False in both Python 2 and Python 3
(4, 5) == [4, 5]
  • prevents from occasional sorting of instances of different types
    sorted([2, '1', 3])  # invalid for Python 3, in Python 2 returns [2, 3, '1']
  • helps to spot some problems that arise when processing raw data

Sidenote: proper check for None is (in both Python versions)

if a is not None:
  pass
  
if a: # WRONG check for None
  pass

Unicode for NLP

s = '您好'
print(len(s))
print(s[:2])

Output:

  • Python 2: 6\n��
  • Python 3: 2\n您好.
x = u'со'
x += 'co' # ok
x += 'со' # fail

Python 2 fails, Python 3 works as expected (because I've used russian letters in strings).

In Python 3 strs are unicode strings, and it is more convenient for NLP processing of non-english texts.

There are other funny things, for instance:

'a' < type < u'a'  # Python 2: True
'a' < u'a'         # Python 2: False
from collections import Counter
Counter('Möbelstück')
  • Python 2: Counter({'\xc3': 2, 'b': 1, 'e': 1, 'c': 1, 'k': 1, 'M': 1, 'l': 1, 's': 1, 't': 1, '\xb6': 1, '\xbc': 1})
  • Python 3: Counter({'M': 1, 'ö': 1, 'b': 1, 'e': 1, 'l': 1, 's': 1, 't': 1, 'ü': 1, 'c': 1, 'k': 1})

You can handle all of this in Python 2 properly, but Python 3 is more friendly.

Preserving order of dictionaries and **kwargs

In CPython 3.6+ dicts behave like OrderedDict by default (and this is guaranteed in Python 3.7+). This preserves order during dict comprehensions (and other operations, e.g. during json serialization/deserialization)

import json
x = {str(i):i for i in range(5)}
json.loads(json.dumps(x))
# Python 2
{u'1': 1, u'0': 0, u'3': 3, u'2': 2, u'4': 4}
# Python 3
{'0': 0, '1': 1, '2': 2, '3': 3, '4': 4}

Same applies to **kwargs (in Python 3.6+), they're kept in the same order as they appear in parameters. Order is crucial when it comes to data pipelines, previously we had to write it in a cumbersome manner:

from torch import nn

# Python 2
model = nn.Sequential(OrderedDict([
          ('conv1', nn.Conv2d(1,20,5)),
          ('relu1', nn.ReLU()),
          ('conv2', nn.Conv2d(20,64,5)),
          ('relu2', nn.ReLU())
        ]))

# Python 3.6+, how it *can* be done, not supported right now in pytorch
model = nn.Sequential(
    conv1=nn.Conv2d(1,20,5),
    relu1=nn.ReLU(),
    conv2=nn.Conv2d(20,64,5),
    relu2=nn.ReLU())
)        

Did you notice? Uniqueness of names is also checked automatically.

Iterable unpacking

# handy when amount of additional stored info may vary between experiments, but the same code can be used in all cases
model_paramteres, optimizer_parameters, *other_params = load(checkpoint_name)

# picking two last values from a sequence
*prev, next_to_last, last = values_history

# This also works with any iterables, so if you have a function that yields e.g. qualities,
# below is a simple way to take only last two values from a list 
*prev, next_to_last, last = iter_train(args)

Default pickle engine provides better compression for arrays

# Python 2
import cPickle as pickle
import numpy
print len(pickle.dumps(numpy.random.normal(size=[1000, 1000])))
# result: 23691675

# Python 3
import pickle
import numpy
len(pickle.dumps(numpy.random.normal(size=[1000, 1000])))
# result: 8000162

Three times less space. And it is much faster. Actually similar compression (but not speed) is achievable with protocol=2 parameter, but users typically ignore this option (or simply not aware of it).

Safer comprehensions

labels = <initial_value>
predictions = [model.predict(data) for data, labels in dataset]

# labels are overwritten in Python 2
# labels are not affected by comprehension in Python 3

Super, simply super()

Python 2 super(...) was a frequent source of mistakes in code.

# Python 2
class MySubClass(MySuperClass):
    def __init__(self, name, **options):
        super(MySubClass, self).__init__(name='subclass', **options)
        
# Python 3
class MySubClass(MySuperClass):
    def __init__(self, name, **options):
        super().__init__(name='subclass', **options)

More on super and method resolution order on stackoverlow.

Better IDE suggestions with variable annotations

The most enjoyable thing about programming in languages like Java, C# and alike is that IDE makes very good suggestions, because type of each identifier is known before executing a program.

In python this is hard to achieve, but annotations will help you

  • write your expectations in a clear form
  • and get good suggestions from IDE


This is an example of PyCharm suggestions with variable annotations. This works even in situations when functions you use are not annotated (e.g. due to backward compatibility).

Multiple unpacking

Here is how you merge two dicts now:

x = dict(a=1, b=2)
y = dict(b=3, d=4)
# Python 3.5+
z = {**x, **y} 
# z = {'a': 1, 'b': 3, 'd': 4}, note that value for `b` is taken from the latter dict.

See this thread at StackOverflow for comparison with Python 2.

The aame approach also works for lists, tuples, and sets (a, b, c are any iterables):

[*a, *b, *c] # list, concatenating 
(*a, *b, *c) # tuple, concatenating 
{*a, *b, *c} # set, union 

Functions also support this for *args and **kwargs:

Python 3.5+
do_something(**{**default_settings, **custom_settings})

# Also possible, this code also checks there is no intersection between keys of dictionaries
do_something(**first_args, **second_args)

Future-proof APIs with keyword-only arguments

Let's consider this snippet

model = sklearn.svm.SVC(2, 'poly', 2, 4, 0.5)

Obviously, an author of this code didn't get Python style of coding yet (most probably, just jumped from cpp or rust). Unfortunately, this is not just question of taste, because changing the order of arguments (adding/deleting) in SVC will break this code. In particular, sklearn does some reordering/renaming from time to time of numerous algorithm parameters to provide consistent API. Each such refactoring may drive to broken code.

In Python 3, library authors may demand explicitly naming parameters by using *:

class SVC(BaseSVC):
    def __init__(self, *, C=1.0, kernel='rbf', degree=3, gamma='auto', coef0=0.0, ... )
  • users have to specify names of parameters sklearn.svm.SVC(C=2, kernel='poly', degree=2, gamma=4, coef0=0.5) now
  • this mechanism provides a great combination of reliability and flexibility of APIs

Minor: constants in math module

# Python 3
math.inf # 'largest' number
math.nan # not a number

max_quality = -math.inf  # no more magic initial values!

for model in trained_models:
    max_quality = max(max_quality, compute_quality(model, data))

Minor: single integer type

Python 2 provides two basic integer types, which are int (64-bit signed integer) and long for long arithmetics (quite confusing after C++).

Python 3 has a single type int, which incorporates long arithmetics.

Here is how you check that value is integer:

isinstance(x, numbers.Integral) # Python 2, the canonical way
isinstance(x, (long, int))      # Python 2
isinstance(x, int)              # Python 3, easier to remember

Other stuff

  • Enums are theoretically useful, but
    • string-typing is already widely adopted in the python data stack
    • Enums don't seem to interplay with numpy and categorical from pandas
  • coroutines also sound very promising for data pipelining (see slides by David Beazley), but I don't see their adoption in the wild.
  • Python 3 has stable ABI
  • Python 3 supports unicode identifies (so ω = Δφ / Δt is ok), but you'd better use good old ASCII names
  • some libraries e.g. jupyterhub (jupyter in cloud), django and fresh ipython only support Python 3, so features that sound useless for you are useful for libraries you'll probably want to use once.

Problems for code migration specific for data science (and how to resolve those)

  • support for nested arguments was dropped

    map(lambda x, (y, z): x, z, dict.items())
    

    However, it is still perfectly working with different comprehensions:

    {x:z for x, (y, z) in d.items()}

    In general, comprehensions are also better 'translatable' between Python 2 and 3.

  • map(), .keys(), .values(), .items(), etc. return iterators, not lists. Main problems with iterators are:

    • no trivial slicing
    • can't be iterated twice

    Almost all of the problems are resolved by converting result to list.

  • see Python FAQ: How do I port to Python 3? when in trouble

Main problems for teaching machine learning and data science with python

Course authors should spend time at first lectures to explain what is an iterator, why it can't be sliced / concatenated / multiplied / iterated twice like a string (and how to deal with it).

I think most of course authors would be happy to avoid these details, but now it is hardly possible.

Conclusion

Python 2 and Python 3 co-exist for almost 10 years, but we should move to Python 3.

Research and production code should become a bit shorter, more readable, and significantly safer after moving to Python 3-only codebase.

Right now most libraries support both Python versions. And I can't wait for the bright moment when packages drop support for Python 2 and enjoy new language features.

Following migrations are promised to be smoother: "we will never do this kind of backwards-incompatible change again"

Links

License

This text was published by Alex Rogozhnikov under CC BY-SA 3.0 License (excluding images).