When creating new numpy arrays, you can make them like this:
a = numpy.array((2, 5))
b = numpy.array((a[0] + 1, 10))
or like this:
a = numpy.array([2, 5])
b = numpy.array([a[0] + 1, 10])
Which way is better?
Add a comment
|
2 Answers
After a few tests based on @leopardshark's answer, it seems that tuples are only better if you initialize the array based on constants. If you initialize from a tuple/list of variables, there's negligible difference
Constants test setup
import dis, timeit
list_timing = timeit.timeit('numpy.array([2, 5])', setup = 'import numpy', number = 1000000)
tuple_timing = timeit.timeit('numpy.array((2, 5))', setup = 'import numpy', number = 1000000)
print(f"List mean time: {list_timing}")
>>> 0.6392972
print(f"Tuple mean time: {tuple_timing}")
>>> 0.6296533
Disassembled outputs here are the same as @leopardshark's
Variables test setup
import dis, timeit
x, y = 2, 5
list_timing = timeit.timeit('numpy.array([x, y])', setup = 'import numpy; x, y = 2, 5', number = 1000000)
tuple_timing = timeit.timeit('numpy.array((x, y))', setup = 'import numpy; x, y = 2, 5', number = 1000000)
print(f"List mean time: {list_timing}")
>>> 0.6279472
print(f"Tuple mean time: {tuple_timing}")
>>> 0.6288363
print(dis.dis('numpy.array([x, y])'))
>>> 1 0 LOAD_NAME 0 (numpy)
>>> 2 LOAD_METHOD 1 (array)
>>> 4 LOAD_NAME 2 (x)
>>> 6 LOAD_NAME 3 (y)
>>> 8 BUILD_LIST 2
>>> 10 CALL_METHOD 1
>>> 12 RETURN_VALUE
print(dis.dis('numpy.array((x, y))'))
>>> 1 0 LOAD_NAME 0 (numpy)
>>> 2 LOAD_METHOD 1 (array)
>>> 4 LOAD_NAME 2 (x)
>>> 6 LOAD_NAME 3 (y)
>>> 8 BUILD_TUPLE 2
>>> 10 CALL_METHOD 1
>>> 12 RETURN_VALUE
The disassembled outputs are the same, while the timings are much closer
Comments
Tuples are about 10% faster.
>>> timeit.timeit("numpy.array((2, 5))", setup="import numpy")
0.9039838570024585
>>> timeit.timeit("numpy.array([2, 5])", setup="import numpy")
1.0044978570003877
I got the same results with the numpy.array((a[0] + 1, 10)) example as well. The dis tool reveals the reason for the difference:
>>> dis.dis("numpy.array((2, 5))")
1 0 LOAD_NAME 0 (numpy)
2 LOAD_METHOD 1 (array)
4 LOAD_CONST 0 ((2, 5))
6 CALL_METHOD 1
8 RETURN_VALUE
>>> dis.dis("numpy.array([2, 5])")
1 0 LOAD_NAME 0 (numpy)
2 LOAD_METHOD 1 (array)
4 LOAD_CONST 0 (2)
6 LOAD_CONST 1 (5)
8 BUILD_LIST 2
10 CALL_METHOD 1
12 RETURN_VALUE
It seems the tuple is treated as a single object as it is created whereas the list needs to be built.
1 Comment
jtlz2
dis - great tip!