[英]Multiply a 3D matrix with a 2D matrix, return a matrix with the same dimensions as the 2D one
In Numpy, say I have a matrix A of dimensions ixjxk, and a matrix B of dimensions kx i.在 Numpy 中,假设我有一个尺寸为 ixjxk 的矩阵 A,和一个尺寸为 kx i 的矩阵 B。 How do I get back a product with dimensions kxi without using a loop?
如何在不使用循环的情况下取回尺寸为 kxi 的产品?
For example:例如:
Let A = [[[1 2],
[3 4]],
[[5 6],
[7 8]]]
B = [[a b],
[c d]]
I would like to get:我想得到:
C = [[a+2c 5b+6d],
[3a+4c 7b+8d]]
My current solution is我目前的解决方案是
np.diagonal(np.dot(A, B), axis1=0, axis2=2)
However, the problem is I'm working with a large data set ( A
and B
have huge dimensions) so np.dot(A, B)
would lead to MemoryError
.但是,问题是我正在处理一个大数据集(
A
和B
有很大的尺寸),所以np.dot(A, B)
会导致MemoryError
。 Therefore I want to figure out a better way to solve this without computing the dot product.因此,我想找出一种更好的方法来解决这个问题,而无需计算点积。
I have looked into functions like einsum
and tensordot
but I didn't find what I need (or maybe I missed something).我已经研究过像
einsum
和tensordot
这样的函数,但我没有找到我需要的东西(或者我可能错过了一些东西)。 I would appreciate if someone could help me out.如果有人可以帮助我,我将不胜感激。 Thank you!
谢谢!
You can do this by analyzing the input and output dimensions.您可以通过分析输入和 output 维度来做到这一点。
A.shape -> i, j, k
B.shape -> k, i
Your example is not very good, but if you look carefully at what you're asking the output to be, the shape of B
must match the first and last dimensions of A
.你的例子不是很好,但如果你仔细看看你要求 output 是什么,
B
的形状必须与A
的第一个和最后一个尺寸相匹配。
The sum reduction is happening along the last axis of A
and the first axis of B
:总和减少发生在
A
的最后一个轴和B
的第一个轴上:
np.einsum('ijk,ki->ji', A, B)
einsum
can be intimidating, but doing this sort of analysis can save you a lot of the frustration caused by trial and error. einsum
可能令人生畏,但进行此类分析可以为您节省大量因反复试验而导致的挫败感。
Looking at your output, C[:,0]
is看着你的 output,
C[:,0]
是
A[0] @ B[0].T
and similarly, C[:,1]
is A[1] @ B[1].T
.同样,
C[:,1]
是A[1] @ B[1].T
。
With that in mind, the np.einsum
formula is:考虑到这一点,
np.einsum
公式为:
C = np.einsum('ijk,ik->ji', A,B)
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.