What is the preferred syntax for initializing a dict: curly brace literals {} or the dict() function?

I'm putting in some effort to learn Python, and I am paying close attention to common coding standards. This may seem like a pointlessly nit-picky question, but I am trying to focus on best-practices as I learn, so I don't have to unlearn any 'bad' habits later.

I see two common methods for initializing a dict:

a = {
    'a': 'value',
    'another': 'value',
}

b = dict( 
    a='value',
    another='value',
)

Which is considered to be "more pythonic"? Which do you use? Why?

Solution 1:

Curly braces. Passing keyword arguments into dict(), though it works beautifully in a lot of scenarios, can only initialize a map if the keys are valid Python identifiers.

This works:

a = {'import': 'trade', 1: 7.8}

a = dict({'import': 'trade', 1: 7.8})

This won't work:

a = dict(import='trade', 1=7.8)

It will result in the following error:

    a = dict(import='trade', 1=7.8)
             ^
SyntaxError: invalid syntax

Solution 2:

The first, curly braces. Otherwise, you run into consistency issues with keys that have odd characters in them, like =.

# Works fine.
a = {
    'a': 'value',
    'b=c': 'value',
}

# Eeep! Breaks if trying to be consistent.
b = dict( 
    a='value',
    b=c='value',
)

Solution 3:

The first version is preferable:

It works for all kinds of keys, so you can, for example, say {1: 'one', 2: 'two'}. The second variant only works for (some) string keys. Using different kinds of syntax depending on the type of the keys would be an unnecessary inconsistency.

It is faster:

$ python -m timeit "dict(a='value', another='value')"
1000000 loops, best of 3: 0.79 usec per loop
$ python -m timeit "{'a': 'value','another': 'value'}"
1000000 loops, best of 3: 0.305 usec per loop

If the special syntax for dictionary literals wasn't intended to be used, it probably wouldn't exist.

Solution 4:

I almost always use curly-braces; however, in some cases where I'm writing tests, I do keyword packing/unpacking, and in these cases dict() is much more maintainable, as I don't need to change:

a=1,
b=2,

to:

'a': 1,
'b': 2,

It also helps in some circumstances where I think I might want to turn it into a namedtuple or class instance at a later time.

In the implementation itself, because of my obsession with optimisation, and when I don't see a particularly huge maintainability benefit, I'll always favour curly-braces.

In tests and the implementation, I would never use dict() if there is a chance that the keys added then, or in the future, would either:

Not always be a string
Not only contain digits, ASCII letters and underscores
Start with an integer (dict(1foo=2) raises a SyntaxError)