Difference between list(numpy_array) and numpy_array.tolist()

Question

What is the difference between applying list() on a numpy array vs. calling tolist()?

I was checking the types of both output

Answer 1

Your example already shows the difference; consider the following 2D array:

>>> import numpy as np
>>> a = np.arange(4).reshape(2, 2)
>>> a
array([[0, 1],
       [2, 3]])
>>> a.tolist()
[[0, 1], [2, 3]] # nested vanilla lists
>>> list(a)
[array([0, 1]), array([2, 3])] # list of arrays

tolist handles the full conversion to nested vanilla lists (i.e. list of list of int), whereas list just iterates over the first dimension of the array, creating a list of arrays (list of np.array of np.int64). Although both are lists:

>>> type(list(a))
<type 'list'>
>>> type(a.tolist())
<type 'list'>

the elements of each list have a different type:

>>> type(list(a)[0])
<type 'numpy.ndarray'>
>>> type(a.tolist()[0])
<type 'list'>

The other difference, as you note, is that list will work on any iterable, whereas tolist can only be called on objects that specifically implement that method.

Answer 2

.tolist() appears to convert all of the values recursively to python primitives (list), whereas list creates a python list from an iterable. Since the numpy array is an array of arrays, list(...) creates a list of arrays

You can think of list as a function that looks like this:

# Not the actually implementation, just for demo purposes
def  list(iterable):
    newlist = []
    for obj in iter(iterable):
        newlist.append(obj)
    return newlist

Answer 3

The major difference is that tolist recursively converts all data to python standard library types.

For instance:

>>> arr = numpy.arange(2)
>>> [type(item) for item in list(arr)]
[numpy.int64, numpy.int64]
>>> [type(item) for item in arr.tolist()]
[builtins.int, builtins.int]

Aside from the functional differences tolist will generally be quicker as it knows it has a numpy array and access to the backing array. Whereas, list will fall back to using an iterator to add all the elements.

In [2]: arr = numpy.arange(1000)

In [3]: %timeit arr.tolist()
10000 loops, best of 3: 33 µs per loop

In [4]: %timeit list(arr)
10000 loops, best of 3: 80.7 µs per loop

I would expect the tolist to be