使用numpy对2d数组进行“平铺”

3 投票
2 回答
2403 浏览
提问于 2025-04-17 11:59

我正在尝试通过将一个二维数组分成多个方块来减小它的大小,然后把这些方块写入另一个数组。每个方块的大小是可变的,比如说边长为n。这个数组的数据类型是整数。目前我在用Python的循环把每个方块分配给一个临时数组,然后从这个临时数组中提取出唯一的值。接着我会遍历这些唯一值,找出出现次数最多的那个。你可以想象,随着输入数组的增大,这个过程会变得非常慢。

我见过一些例子是从这些方块中取最小值、最大值和平均值,但我不知道怎么把它们转换成多数值。对二维numpy数组进行平均分组用平均值调整大小或重新分组一个numpy二维数组

我希望能找到一些方法来加快这个过程,使用numpy来对整个数组进行处理。(当输入数据太大而无法放入内存时,我可以处理这个方面,切换到数组的分块部分)

谢谢

#snippet of my code
#pull a tmpArray representing one square chunk of my input array
kernel = sourceDs.GetRasterBand(1).ReadAsArray(int(sourceRow), 
                                    int(sourceCol), 
                                    int(numSourcePerTarget),
                                    int(numSourcePerTarget))
#get a list of the unique values
uniques = np.unique(kernel)
curMajority = -3.40282346639e+038
for val in uniques:
    numOccurances = (array(kernel)==val).sum()
    if numOccurances > curMajority:
        ans = val
        curMajority = numOccurances

#write out our answer
outBand.WriteArray(curMajority, row, col)

#This is insanity!!!

在Bago的优秀建议下,我觉得我已经接近解决方案了。到目前为止,我做了一些改动,使用了一个(xy, nn)数组,来自原始网格的形状。现在我遇到的问题是,我似乎无法找到如何将一维的where、counts和uniq_a步骤转换为二维的方式。

#test data
grid = np.array([[ 37,  1,  4,  4, 6,  6,  7,  7],
                 [ 1,  37,  4,  5, 6,  7,  7,  8],
                 [ 9,  9, 11, 11, 13,  13,  15,  15],
                 [9, 10, 11, 12, 13,  14,  15,  16],
                 [ 17, 17,  19,  19, 21,  11,  23,  23],
                 [ 17, 18,  19,  20, 11,  22,  23,  24],
                 [ 25, 25, 27, 27, 29,  29,  31,  32],
                 [25, 26, 27, 28, 29,  30,  31,  32]])
print grid

n = 4
X, Y = grid.shape
x = X // n
y = Y // n
grid = grid.reshape( (x, n, y, n) )
grid = grid.transpose( [0, 2, 1, 3] )
grid = grid.reshape( (x*y, n*n) )
grid = np.sort(grid)
diff = np.empty((grid.shape[0], grid.shape[1]+1), bool)
diff[:, 0] = True
diff[:, -1] = True
diff[:, 1:-1] = grid[:, 1:] != grid[:, :-1]
where = np.where(diff)

#This is where if falls apart for me as 
#where returns two arrays:
# row indices [0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 3 3 3 3 3 3 3 3 3 3]
# col indices [ 0  2  5  6  9 10 13 14 16  0  3  7  8 11 12 15 16  0  3  4  7  8 11 12 15
# 16  0  2  3  4  7  8 11 12 14 16]
#I'm not sure how to get a 
counts = where[:, 1:] - where[:, -1]
argmax = counts[:].argmax()
uniq_a = grid[diff[1:]]
print uniq_a[argmax]

2 个回答

1

这可能有点偷懒,但我最后还是用了scipy.stats.stats库里的mode函数来找出出现次数最多的值。我不太确定这个方法在处理时间上和其他解决方案相比怎么样。

import scipy.stats.stats as stats
#test data
grid = np.array([[ 37,  1,  4,  4, 6,  6,  7,  7],
                 [ 1,  37,  4,  5, 6,  7,  7,  8],
                 [ 9,  9, 11, 11, 13,  13,  15,  15],
                 [9, 10, 11, 12, 13,  14,  15,  16],
                 [ 17, 17,  19,  19, 21,  11,  23,  23],
                 [ 17, 18,  19,  20, 11,  22,  23,  24],
                 [ 25, 25, 27, 27, 29,  29,  31,  32],
                 [25, 26, 27, 28, 29,  30,  31,  32]])
print grid

n = 2
X, Y = grid.shape
x = X // n
y = Y // n
grid = grid.reshape( (x, n, y, n) )
grid = grid.transpose( [0, 2, 1, 3] )
grid = grid.reshape( (x*y, n*n) )
answer =  np.array(stats.mode(grid, 1)[0]).reshape(x, y)
3

这里有一个函数,可以更快地找到大多数元素,它是基于numpy.unique的实现。

def get_majority(a):
    a = a.ravel()
    a = np.sort(a)
    diff = np.empty(len(a)+1, 'bool')
    diff[0] = True
    diff[-1] = True
    diff[1:-1] = a[1:] != a[:-1]
    where = np.where(diff)[0]
    counts = where[1:] - where[:-1]
    argmax = counts.argmax()
    uniq_a = a[diff[1:]]
    return uniq_a[argmax]

如果这对你有帮助,请告诉我。

更新

你可以这样做,把你的数组变成 (n*n, x, y),这样就可以在第一个轴上操作,以更高效的方式完成任务。

X, Y = a.shape
x = X // n
y = Y // n
a = a.reshape( (x, n, y, n) )
a = a.transpose( [1, 3, 0, 2] )
a = a.reshape( (n*n, x, y) )

有几点需要注意。虽然reshape和transpose在可能的情况下会返回视图,但我认为reshape-转置-reshape会被迫复制。此外,将上述方法推广到其他轴上操作应该是可行的,但可能需要一些创造力。

撰写回答