NumPy functions are designed to work with NumPy classes (e.g. numpy.array), whilst many work with core Python classes (e.g. list) this normally incurs an additional cost of NumPy first converting the argument to the corresponding NumPy type.

If you are passing long list this cost is higher, and if you’re performing it repeatedly you’re incurring that redundant conversion multiple times.

Example Code

The below example sees an identical list and numpy. array constructed of 1 million length.

numpy.random.choice() is then called to select a random element 100 times.

import timeit

# Common untimed setup
setup = f"""
N = 1000000
import numpy as np
ls = list(range(1000000))
ar = np.arange(N)
"""

# List version
ls_t = timeit.timeit('np.random.choice(ls)', number=100, setup=setup)
 
# NumPy array version
ar_t = timeit.timeit('np.random.choice(ar)', number=100, setup=setup)

print("np.random.choice(list): %.4fms"%(ls_t))
print("np.random.choice(np.array): %.4fms"%(ar_t))

Passing a numpy.array was over 4000x faster!

np.random.choice(list): 4.3459ms
np.random.choice(np.array): 0.0010ms

So where possible, use a numpy.array, even if that means temporarily converting the list yourself before making repeated calls to NumPy functions.

The Technical Detail

When this issue was first identified, it was shown that numpy.random.choice() always calls np.array(..., copy=False) when passed a list (source link).

We can benchmark that too

import timeit

# Common untimed setup
setup = f"""
N = 1000000
import numpy as np
ls = list(range(1000000))
"""

# List time
ls_t = timeit.timeit('np.random.choice(ls)', number=100, setup=setup)
 
# List copy time
lc_t = timeit.timeit('np.array(ls, copy=False)', number=100, setup=setup)

print("np.random.choice(list): %.2fms"%(ls_t))
print("np.array(list, copy=False): %.2fms"%(lc_t))

Which shows that 99% of the runtime is the copy.

np.random.choice(list): 4.37ms
np.array(list, copy=False): 4.32ms

NumPy developers confirmed in the bug report that this performance issue is more widespread, and not something they intend to resolve.

For the example used in this guide, generating a random integer to select the element (e.g. numpy.random.randint(len(ls))) would be even faster yet. This is regardless of whether a list or numpy.array is used, as it avoids some additional branching due to the complex configuration numpy.random.choice() permits.