Trois options:
- Quelque chose de semblable à la liste en utilisant la compréhension de
.item
NumPy.
itertools.starmap
np.broadcast
Le timing est ci-dessous. L'option 3 semble être de loin la voie la plus rapide.
from itertools import starmap
np.random.seed(123)
target = np.random.randn(2, 2)
roundto = np.arange(1, 5, dtype=np.int16).reshape((2, 2)) # must be type int
def method1():
return (np.array([round(target.item(i), roundto.item(j))
for i, j in zip(range(target.size),
range(roundto.size))])
.reshape(target.shape))
def method2():
return np.array(list(starmap(round, zip(target.flatten(),
roundto.flatten())))).reshape(target.shape)
def method3():
b = np.broadcast(target, roundto)
out = np.empty(b.shape)
out.flat = [round(u,v) for (u,v) in b]
return out
from timeit import timeit
timeit(method1, number=100)
Out[50]: 0.003252145578553467
timeit(method2, number=100)
Out[51]: 0.002063405777064986
timeit(method3, number=100)
Out[52]: 0.0009481473990007316
print('method 3 is %0.2f x faster than method 2' %
(timeit(method2, number=100)/timeit(method3, number=100)))
method 3 is 2.91 x faster than method 2
Il ne semble pas logique d'arrondir 24.324 à 2. – Psidom
Je viens de tirer un exemple de ma manche, c'est pourquoi je n'ai pas remarqué cela. –