将 numpy 数组的组名映射到索引的最快方法是什么？

Question

我正在使用激光雷达的 3D 点云。这些点由 numpy 数组给出，如下所示：

points = np.array([[61651921, 416326074, 39805], [61605255, 416360555, 41124], [61664810, 416313743, 39900], [61664837, 416313749, 39910], [61674456, 416316663, 39503], [61651933, 416326074, 39802], [61679969, 416318049, 39500], [61674494, 416316677, 39508], [61651908, 416326079, 39800], [61651908, 416326087, 39802], [61664845, 416313738, 39913], [61674480, 416316668, 39503], [61679996, 416318047, 39510], [61605290, 416360572, 41118], [61605270, 416360565, 41122], [61683939, 416313004, 41052], [61683936, 416313033, 41060], [61679976, 416318044, 39509], [61605279, 416360555, 41109], [61664837, 416313739, 39915], [61674487, 416316666, 39505], [61679961, 416318035, 39503], [61683943, 416313004, 41054], [61683930, 416313042, 41059]])

我想将我的数据分组到大小的立方体中，50*50*50以便每个立方体都保留一些points它包含的可散列索引和 numpy 索引。为了进行拆分，我将cubes = points \\ 50哪些输出分配给：

cubes = np.array([[1233038, 8326521, 796], [1232105, 8327211, 822], [1233296, 8326274, 798], [1233296, 8326274, 798], [1233489, 8326333, 790], [1233038, 8326521, 796], [1233599, 8326360, 790], [1233489, 8326333, 790], [1233038, 8326521, 796], [1233038, 8326521, 796], [1233296, 8326274, 798], [1233489, 8326333, 790], [1233599, 8326360, 790], [1232105, 8327211, 822], [1232105, 8327211, 822], [1233678, 8326260, 821], [1233678, 8326260, 821], [1233599, 8326360, 790], [1232105, 8327211, 822], [1233296, 8326274, 798], [1233489, 8326333, 790], [1233599, 8326360, 790], [1233678, 8326260, 821], [1233678, 8326260, 821]])

我想要的输出如下所示：

{(1232105, 8327211, 822): [1, 13, 14, 18]), 
(1233038, 8326521, 796): [0, 5, 8, 9], 
(1233296, 8326274, 798): [2, 3, 10, 19], 
(1233489, 8326333, 790): [4, 7, 11, 20], 
(1233599, 8326360, 790): [6, 12, 17, 21], 
(1233678, 8326260, 821): [15, 16, 22, 23]}

我的真实点云包含多达几亿个 3D 点。进行这种分组的最快方法是什么？

我已经尝试了大多数各种解决方案。这是假设点的大小约为 2000 万且不同立方体的大小约为 100 万的时间消耗比较：

熊猫 [元组（元素）-> np.array(dtype=int64)]

import pandas as pd
print(pd.DataFrame(cubes).groupby([0,1,2]).indices)
#takes 9sec

默认 [elem.tobytes() 或元组 -> 列表]

#thanks @abc:
result = defaultdict(list)
for idx, elem in enumerate(cubes):
    result[elem.tobytes()].append(idx) # takes 20.5sec
    # result[elem[0], elem[1], elem[2]].append(idx) #takes 27sec
    # result[tuple(elem)].append(idx) # takes 50sec

numpy_indexed [int -> np.array]

# thanks @Eelco Hoogendoorn for his library
values = npi.group_by(cubes).split(np.arange(len(cubes)))
result = dict(enumerate(values))
# takes 9.8sec

Pandas + 降维 [int -> np.array(dtype=int64)]

# thanks @Divakar for showing numexpr library:
import numexpr as ne
def dimensionality_reduction(cubes):
    #cubes = cubes - np.min(cubes, axis=0) #in case some coords are negative 
    cubes = cubes.astype(np.int64)
    s0, s1 = cubes[:,0].max()+1, cubes[:,1].max()+1
    d = {'s0':s0,'s1':s1,'c0':cubes[:,0],'c1':cubes[:,1],'c2':cubes[:,2]}
    c1D = ne.evaluate('c0+c1*s0+c2*s0*s1',d)
    return c1D
cubes = dimensionality_reduction(cubes)
result = pd.DataFrame(cubes).groupby([0]).indices
# takes 2.5 seconds

可以在这里下载cubes.npz文件并使用命令

cubes = np.load('cubes.npz')['array']

检查性能时间。

Answer 1

每组的索引数量恒定

方法#1

我们可以执行dimensionality-reduction以减少cubes到一维数组。这是基于将给定的立方体数据映射到 n-dim 网格以计算线性索引等价物，详细讨论here。然后，基于这些线性索引的唯一性，我们可以分离唯一组及其对应的索引。因此，按照这些策略，我们会有一个解决方案，就像这样——

N = 4 # number of indices per group
c1D = np.ravel_multi_index(cubes.T, cubes.max(0)+1)
sidx = c1D.argsort()
indices = sidx.reshape(-1,N)
unq_groups = cubes[indices[:,0]]

# If you need in a zipped dictionary format
out = dict(zip(map(tuple,unq_groups), indices))

备选方案#1：如果整数值cubes太大，我们可能希望dimensionality-reduction选择范围较短的维度作为主轴。因此，对于这些情况，我们可以修改缩减步骤以获得c1D，就像这样 -

s1,s2 = cubes[:,:2].max(0)+1
s = np.r_[s2,1,s1*s2]
c1D = cubes.dot(s)

方法#2

接下来，我们可以使用Cython-powered kd-tree快速最近邻查找来获得最近的相邻索引，从而像这样解决我们的情况 -

from scipy.spatial import cKDTree

idx = cKDTree(cubes).query(cubes, k=N)[1] # N = 4 as discussed earlier
I = idx[:,0].argsort().reshape(-1,N)[:,0]
unq_groups,indices = cubes[I],idx[I]

一般情况：每组可变数量的索引

我们将通过一些拆分扩展基于 argsort 的方法以获得我们想要的输出，就像这样 -

c1D = np.ravel_multi_index(cubes.T, cubes.max(0)+1)

sidx = c1D.argsort()
c1Ds = c1D[sidx]
split_idx = np.flatnonzero(np.r_[True,c1Ds[:-1]!=c1Ds[1:],True])
grps = cubes[sidx[split_idx[:-1]]]

indices = [sidx[i:j] for (i,j) in zip(split_idx[:-1],split_idx[1:])]
# If needed as dict o/p
out = dict(zip(map(tuple,grps), indices))

使用 1D 版本的组cubes作为键

我们将使用 group of cubesas 键扩展前面列出的方法，以简化字典创建过程并使其高效，就像这样 -

def numpy1(cubes):
    c1D = np.ravel_multi_index(cubes.T, cubes.max(0)+1)        
    sidx = c1D.argsort()
    c1Ds = c1D[sidx]
    mask = np.r_[True,c1Ds[:-1]!=c1Ds[1:],True]
    split_idx = np.flatnonzero(mask)
    indices = [sidx[i:j] for (i,j) in zip(split_idx[:-1],split_idx[1:])]
    out = dict(zip(c1Ds[mask[:-1]],indices))
    return out

接下来，我们将使用numba包来迭代并获得最终的可哈希字典输出。随之而来，将有两种解决方案 - 一种分别使用获取键和值numba，主调用将压缩并转换为 dict，而另一种将创建一个numba-supporteddict 类型，因此主调用函数不需要额外的工作.

因此，我们将有第一个numba解决方案：

from numba import  njit

@njit
def _numba1(sidx, c1D):
    out = []
    n = len(sidx)
    start = 0
    grpID = []
    for i in range(1,n):
        if c1D[sidx[i]]!=c1D[sidx[i-1]]:
            out.append(sidx[start:i])
            grpID.append(c1D[sidx[start]])
            start = i
    out.append(sidx[start:])
    grpID.append(c1D[sidx[start]])
    return grpID,out

def numba1(cubes):
    c1D = np.ravel_multi_index(cubes.T, cubes.max(0)+1)
    sidx = c1D.argsort()
    out = dict(zip(*_numba1(sidx, c1D)))
    return out

第二种numba解决方案为：

from numba import types
from numba.typed import Dict

int_array = types.int64[:]

@njit
def _numba2(sidx, c1D):
    n = len(sidx)
    start = 0
    outt = Dict.empty(
        key_type=types.int64,
        value_type=int_array,
    )
    for i in range(1,n):
        if c1D[sidx[i]]!=c1D[sidx[i-1]]:
            outt[c1D[sidx[start]]] = sidx[start:i]
            start = i
    outt[c1D[sidx[start]]] = sidx[start:]
    return outt

def numba2(cubes):
    c1D = np.ravel_multi_index(cubes.T, cubes.max(0)+1)    
    sidx = c1D.argsort()
    out = _numba2(sidx, c1D)
    return out

cubes.npz数据时间-

In [4]: cubes = np.load('cubes.npz')['array']

In [5]: %timeit numpy1(cubes)
   ...: %timeit numba1(cubes)
   ...: %timeit numba2(cubes)
2.38 s ± 14.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
2.13 s ± 25.2 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
1.8 s ± 5.95 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

替代方案#1：我们可以通过numexpr大型数组的计算实现进一步的加速c1D，就像这样 -

import numexpr as ne

s0,s1 = cubes[:,0].max()+1,cubes[:,1].max()+1
d = {'s0':s0,'s1':s1,'c0':cubes[:,0],'c1':cubes[:,1],'c2':cubes[:,2]}
c1D = ne.evaluate('c0+c1*s0+c2*s0*s1',d)

这将适用于所有需要c1D.

Answer 2

您可能只是迭代并将每个元素的索引添加到相应的列表中。

from collections import defaultdict

res = defaultdict(list)

for idx, elem in enumerate(cubes):
    #res[tuple(elem)].append(idx)
    res[elem.tobytes()].append(idx)

通过使用tobytes()而不是将键转换为元组，可以进一步改进运行时。