Python copy.copy() vs NumPy np.copy()

Question

Is there any situation where I would want to use NumPy's np.copy() over Python's copy.copy() method? As far as I can tell, both create shallow copies, but NumPy is limited to arrays. Is there any performance or special handling by NumPy that's not accounted for by copy?

• cs95: For starters, the implementation would differ. Furthermore, you don't have to import copy to use copy.
• hpaulj: We can double check this, but I think copy.copy just delegates the action to the objects own copy method, where possible. In fact you can read the copy.copy code yourself. arr.__copy__ is a method. np.copy is also python code.
• hpaulj: I see and use arr.copy() most often.
• hpaulj: deepcopy has an added memo parameter, which is discussed at Python: Numpy deepcopy TypeError

Answer 1

Yes, there are side effects and numpy code is around 20% faster than pure python for float64 types.

In [1]: import numpy as np
In [2]: from copy import copy
In [3]: arr = np.random.rand(10000, 10000)
In [4]: %timeit copy(arr)
535 ms ± 97.8 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [5]: %timeit np.copy(arr)
453 ms ± 19.3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
In [6]: %timeit arr.copy()
456 ms ± 22 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Answer 2

numpy.copy allows more control over the memory layout of the result with the order argument, and it always produces an array, even if given some other array-like. Also, you don't have to separately import the copy module.

• dkv: But, does importing copy have any side-effects beyond the extra line of code?