在NumPy数组中计算连续的1

这是向量化方法-

def island_cumsum_vectorized(a):
    a_ext = np.concatenate(( [0], a, [0] ))
    idx = np.flatnonzero(a_ext[1:] != a_ext[:-1])
    a_ext[1:][idx[1::2]] = idx[::2] - idx[1::2]
    return a_ext.cumsum()[1:-1]

样品运行-

In [91]: a = np.array([1, 1, 1, 0, 0, 0, 1, 1, 0, 0])

In [92]: island_cumsum_vectorized(a)
Out[92]: array([1, 2, 3, 0, 0, 0, 1, 2, 0, 0])

In [93]: a = np.array([0, 1, 1, 1, 1, 0, 0, 0, 1, 1, 0, 0, 1])

In [94]: island_cumsum_vectorized(a)
Out[94]: array([0, 1, 2, 3, 4, 0, 0, 0, 1, 2, 0, 0, 1])

运行时测试

对于时间，我将使用OP的示例输入数组并重复/平铺它，希望这应该是less opportunistic benchmark-

小箱：

In [16]: a = np.array([1, 1, 1, 0, 0, 0, 1, 1, 0, 0])

In [17]: a = np.tile(a,10)  # Repeat OP's data 10 times

# @Paul Panzer's solution
In [18]: %timeit np.concatenate([np.cumsum(c) if c[0] == 1 else c for c in np.split(a, 1 + np.where(np.diff(a))[0])])
10000 loops, best of 3: 73.4 µs per loop

In [19]: %timeit island_cumsum_vectorized(a)
100000 loops, best of 3: 8.65 µs per loop

更大的情况：

In [20]: a = np.array([1, 1, 1, 0, 0, 0, 1, 1, 0, 0])

In [21]: a = np.tile(a,1000)  # Repeat OP's data 1000 times

# @Paul Panzer's solution
In [22]: %timeit np.concatenate([np.cumsum(c) if c[0] == 1 else c for c in np.split(a, 1 + np.where(np.diff(a))[0])])
100 loops, best of 3: 6.52 ms per loop

In [23]: %timeit island_cumsum_vectorized(a)
10000 loops, best of 3: 49.7 µs per loop

不，我想要非常大的案件：

In [24]: a = np.array([1, 1, 1, 0, 0, 0, 1, 1, 0, 0])

In [25]: a = np.tile(a,100000)  # Repeat OP's data 100000 times

# @Paul Panzer's solution
In [26]: %timeit np.concatenate([np.cumsum(c) if c[0] == 1 else c for c in np.split(a, 1 + np.where(np.diff(a))[0])])
1 loops, best of 3: 725 ms per loop

In [27]: %timeit island_cumsum_vectorized(a)
100 loops, best of 3: 7.28 ms per loop