给定索引和大小,是否有更有效的方法来生成标准基向量:
import numpy as np
np.array([1.0 if i == index else 0.0 for i in range(size)])
In [2]: import numpy as np
In [9]: size = 5
In [10]: index = 2
In [11]: np.eye(1,size,index)
Out[11]: array([[ 0., 0., 1., 0., 0.]])
嗯,不幸的是,为此使用 np.eye 相当慢:
In [12]: %timeit np.eye(1,size,index)
100000 loops, best of 3: 7.68 us per loop
In [13]: %timeit a = np.zeros(size); a[index] = 1.0
1000000 loops, best of 3: 1.53 us per loop
包装np.zeros(size); a[index] = 1.0
在一个函数中只会产生适度的差异,并且仍然比np.eye
:
In [24]: def f(size, index):
....: arr = np.zeros(size)
....: arr[index] = 1.0
....: return arr
....:
In [27]: %timeit f(size, index)
1000000 loops, best of 3: 1.79 us per loop
x = np.zeros(size)
x[index] = 1.0
至少我认为是这样...
>>> t = timeit.Timer('np.array([1.0 if i == index else 0.0 for i in range(size)]
)','import numpy as np;size=10000;index=5123')
>>> t.timeit(10)
0.039461429317952934 #original method
>>> t = timeit.Timer('x=np.zeros(size);x[index]=1.0','import numpy as np;size=10000;index=5123')
>>> t.timeit(10)
9.4077963240124518e-05 #zeros method
>>> t = timeit.Timer('x=np.eye(1.0,size,index)','import numpy as np;size=10000;index=5123')
>>> t.timeit(10)
0.0001398340635319073 #eye method
看起来 np.zeros 是最快的......
我不确定这是否更快,但对我来说肯定更清楚。
a = np.zeros(size)
a[index] = 1.0
通常,您需要的不是一个,而是所有的基向量。如果是这种情况,请考虑np.eye
:
basis = np.eye(3)
for vector in basis:
...
不完全相同,但密切相关:这甚至可以通过一些技巧获得一组基矩阵:
>>> d, e = 2, 3 # want 2x3 matrices
>>> basis = np.eye(d*e,d*e).reshape((d*e,d,e))
>>> print(basis)
[[[ 1. 0. 0.]
[ 0. 0. 0.]]
[[ 0. 1. 0.]
[ 0. 0. 0.]]
[[ 0. 0. 1.]
[ 0. 0. 0.]]
[[ 0. 0. 0.]
[ 1. 0. 0.]]
[[ 0. 0. 0.]
[ 0. 1. 0.]]
[[ 0. 0. 0.]
[ 0. 0. 1.]]]
等等。
它可能不是最快的,但该方法scipy.signal.unit_impulse
将上述概念推广到任何形状的 numpy 数组。
另一种实现方式是:
>>> def f(大小,索引): ... return (np.arange(size) == index).astype(float) ...
这会导致执行时间稍慢:
>>> timeit.timeit('f(size, index)', 'from __main__ import f, size, index', number=1000000) 2.2554846050043125