questions about feature_dimension in hadamard_utils.get_hadK()

[HaoKun-Li]

I notice that hadamard_utils.get_hadK() will check the dimension of n, and return exactly hadamard matrix for specifically n.
However, if I have n(such as 10944) that do not satisfy the if/else judge(10944/64=171, 171 is not the pow of 2), how can I get the appropriate hadamard matrix for n = 10944?

Looking forward to your reply.

[sashkboos]

Thanks @HaoKun-Li for your issue.

Not all sizes have Hadamard matrices. We know those matrices for specific sizes and finding a new Hadamard matrix is a research question. You can find the list of all known Hadamard matrices here: http://neilsloane.com/hadamard/

Note that the get_hadk matrix just returns the above matrices (in the form of Pytorch Tensors). In any case, I think we don't have any odd Hadamard matrix.

However, what I can suggest for your usecase is to apply grouped-hadamard in your model. For example, if you have a matrix of size bx10944 (for some b), you can create a 10944x10944 matrix where you put Hadamards of size 64 on the diagonal and then multiply that to your initial matrix. This is equivalent to apply Hadamard matrices of 64 into every group of 64 columns in your matrix.