Memory Implementation using SVD

[MichaelOVertolli]

This is the memory task where a vector is multiplied by the weight matrix of base components for a set of images and the image intensity vector is output as a result

[cmekik]

using tags... fancy.

[cmekik]

@MichaelOVertolli I am actually working on a utils function that has the following signature:

SVD_compress(source, to_compress) -> compressed

where source, is a matrix which will be used in determining the SVD compression, to_compress is a list of matrices to be compressed with the SVD compression resulting from taking the SVD of source.

So if you want to do SVD compression the way we were doing it in the old models you would write

cpd_stimulus, cpd_encs, cpd_eval_pts = SVD_compress(encs, [stimulus, encs, eval_pts])

I think it will be useful to you.

[MichaelOVertolli]

You're going to have to explain this more. Also, I found some neat tricks in "Programming Computer Vision in Python" (p. 27) that speed up the algorithm if the dimensionality is high (e.g., using the covariance matrix of XX^T).

[cmekik]

It is already written and should be in utils.init. It's called SVDcompressor and is actually a class. Check the documentation, if anything is mysterious. I'd also be glad to help, if things still aren't clear.