Numpy arrays vs Python arrays [duplicate]

Question

I noticed that the de facto standard for array manipulation in Python is through the excellent numpy library. However, I know that the Python Standard Library has an array module, which seems to me to have a similar use-case as Numpy.

Is there any actual real-world example where array is desirable over numpy or just plain list?

From my naive interpretation, array is just memory-efficient container for homogeneous data, but offers no means of improving computational efficiency.

---

EDIT

Just out of curiosity, I searched through Github and import array for Python hits 186'721 counts, while import numpy hits 8'062'678 counts.

However, I could not find a popular repository using array.

Answer 1

To understand the differences between numpy and array, I ran a few more quantitative test.

What I have found is that, for my system (Ubuntu 18.04, Python3), array seems to be twice as fast at generating a large array from the range generator compared to numpy (although numpy's dedicated np.arange() seems to be much faster -- actually too fast, and perhaps it is caching something during tests), but twice as slow than using list.

However, quite surprisingly, array objects seems to be larger than the numpy counterparts. Instead, the list objects are roughly 8-13% larger than array objects (this will vary with the size of the individual items, obviously). Compared to list, array offers a way to control the size of the number objects.

So, perhaps, the only sensible use case for array is actually when numpy is not available.

For completeness, here is the code that I used for the tests:

import numpy as np
import array
import sys

num = int(1e6)
num_i = 100
x = np.logspace(1, int(np.log10(num)), num_i).astype(int)

%timeit list(range(num))
# 10 loops, best of 3: 32.8 ms per loop

%timeit array.array('l', range(num))
# 10 loops, best of 3: 86.3 ms per loop

%timeit np.array(range(num), dtype=np.int64)
# 10 loops, best of 3: 180 ms per loop

%timeit np.arange(num, dtype=np.int64)
# 1000 loops, best of 3: 809 µs per loop


y_list = np.array([sys.getsizeof(list(range(x_i))) for x_i in x])
y_array = np.array([sys.getsizeof(array.array('l', range(x_i))) for x_i in x])
y_np = np.array([sys.getsizeof(np.array(range(x_i), dtype=np.int64)) for x_i in x])

import matplotlib.pyplot as plt

plt.figure(figsize=(12, 6))
plt.plot(x, y_list, label='list')
plt.plot(x, y_array, label='array')
plt.plot(x, y_np, label='numpy')
plt.legend()
plt.show()