在numpy或pytorch中自动获取对角矩阵条纹 [英] Getting diagonal matrix stripe automatically in numpy or pytorch
问题描述
我需要获得矩阵的对角线条纹(不确定此处的术语,对角线矩阵条纹似乎最能描述它).
I need to get a diagonal stripe of the matrix (not sure about the terminology here, diagonal matrix stripe seems to describe it best).
说,我有一个大小为KxN的矩阵,其中K和N是任意大小,K> N.说,我有一个矩阵:
Say, I have a matrix of size KxN, where K and N are arbitrary sizes and K>N. Say, I have a matrix:
[[ 0 1 2]
[ 3 4 5]
[ 6 7 8]
[ 9 10 11]]
我需要从中提取对角线条纹,在这种情况下,是通过截断原始像素而创建的矩阵MxV大小:
From it I would need to extract a diagonal stripe, in this case, a matrix MxV size that is created by truncating the original one:
[[ 0 x x]
[ 3 4 x]
[ x 7 8]
[ x x 11]]
所以结果矩阵为:
[[ 0 4 8]
[ 3 7 11]]
这是一个使用矩阵蒙版的小示例代码,以去除蒙版位置:
Here is a small example code using masking for the matrices, to strip out the masked positions:
import numpy as np
X=np.arange(12).reshape(4,3)
mask=np.asarray([
[ True, False, False],
[ True, True, False],
[ False, True, True],
[ False, False, True]
])
>>> mask
array([[ True, False, False],
[ True, True, False],
[False, True, True],
[False, False, True]], dtype=bool)
>>> X
array([[ 0, 1, 2],
[ 3, 4, 5],
[ 6, 7, 8],
[ 9, 10, 11]])
>>> X.T[mask.T].reshape(3,2).T
array([[ 0, 4, 8],
[ 3, 7, 11]])
但是我看不到如何自动将这种蒙版生成为任何K和N尺寸. 39x9或360x96
But I don't see how such a mask could be automatically generated to any K and N sizes, e.i. 39x9 or 360x96
感谢您的帮助.也许有一些功能可以自动在numpy,scipy或pytorch中执行此操作?
Any help is appreciated. Maybe there is some function that does this automatically either in numpy, scipy or pytorch?
我还有一个问题,是否可能代替:
I’ve got another question, is it possible instead of getting:
[[ 0 x x]
[ 3 4 x]
[ x 7 8]
[ x x 11]]
要获得像这样的反向条纹:
To get a reverse stripe like this:
[[ x x 2]
[ x 4 5]
[ 6 7 x]
[ 9 x x]]
推荐答案
stride_tricks
做到了:
>>> import numpy as np
>>>
>>> def stripe(a):
... a = np.asanyarray(a)
... *sh, i, j = a.shape
... assert i >= j
... *st, k, m = a.strides
... return np.lib.stride_tricks.as_strided(a, (*sh, i-j+1, j), (*st, k, k+m))
...
>>> a = np.arange(24).reshape(6, 4)
>>> a
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11],
[12, 13, 14, 15],
[16, 17, 18, 19],
[20, 21, 22, 23]])
>>> stripe(a)
array([[ 0, 5, 10, 15],
[ 4, 9, 14, 19],
[ 8, 13, 18, 23]])
如果a
是一个数组,则会创建一个可写的视图,这意味着如果您觉得自己很愿意,可以做类似的事情
If a
is an array this creates a writable view, meaning that if you feel so inclined you can do things like
>>> stripe(a)[...] *= 10
>>> a
array([[ 0, 1, 2, 3],
[ 40, 50, 6, 7],
[ 80, 90, 100, 11],
[ 12, 130, 140, 150],
[ 16, 17, 180, 190],
[ 20, 21, 22, 230]])
更新:可以相同的方式获得从左下到右上的条纹.只是轻微的复杂性:它与原始数组的地址不同.
UPDATE: bottom-left to top-right stripes can be obtained in the same spirit. Only minor complication: It is not based at the same address as the original array.
>>> def reverse_stripe(a):
... a = np.asanyarray(a)
... *sh, i, j = a.shape
... assert i >= j
... *st, k, m = a.strides
... return np.lib.stride_tricks.as_strided(a[..., j-1:, :], (*sh, i-j+1, j), (*st, k, m-k))
...
>>> a = np.arange(24).reshape(6, 4)
>>> reverse_stripe(a)
array([[12, 9, 6, 3],
[16, 13, 10, 7],
[20, 17, 14, 11]])
这篇关于在numpy或pytorch中自动获取对角矩阵条纹的文章就介绍到这了,希望我们推荐的答案对大家有所帮助,也希望大家多多支持IT屋!