Python 中嵌套 for 循环的向量化

Question

我有以下嵌套 for 循环（为简单起见随机）：

import numpy as np 

lat_idx = np.random.randint(121, size = 4800)
lon_idx = np.random.randint(201, size = (4800,4800))
sum_cell = np.zeros((121,201))
data = np.random.rand(4800,4800)
for j in range(4800):
    for i in range(4800):
        if lat_idx[i] < 0 or lon_idx[i, j] < 0: 
            continue
        sum_cell[lat_idx[i], lon_idx[i, j]] += data[i, j]

#print(sum_cell)

有没有办法把它写成矩阵运算或者用一些"numpy action"？目前它真的很慢。我的问题是 lon_idx 既依赖于 i 又依赖于 j.

Answer 1

这就是你如何以矢量化的方式做到这一点：

import numpy as np

# Make input data
np.random.seed(0)
data = np.random.rand(4800, 4800)
# Add some negative values in indices
lat_idx = np.random.randint(-20, 121, size=4800)
lon_idx = np.random.randint(-50, 201, size=(4800, 4800))
# Output array
sum_cell = np.zeros((121, 201))
# Make mask for positive indices
lat_idx2 = lat_idx[:, np.newaxis]
m = (lat_idx2 >= 0) & (lon_idx >= 0)
# Get positive indices
lat_pos, lon_pos = np.broadcast_to(lat_idx2, m.shape)[m], lon_idx[m]
# Add values
np.add.at(sum_cell, (lat_pos, lon_pos), data[m])
# Check result with previous method
sum_cell2 = np.zeros((121, 201))
for j in range(4800):
    for i in range(4800):
        if lat_idx[i] < 0 or lon_idx[i, j] < 0: 
            continue
        sum_cell2[lat_idx[i], lon_idx[i, j]] += data[i, j]
print(np.allclose(sum_cell, sum_cell2))
# True

Python 中嵌套 for 循环的向量化

Vectorization of nested for loop in Python

python

for-loop

numpy

vectorization