Derivation of matrices G and B

[buttercutter]

Could anyone advise on how to derive matrices G and B from matrix A using Fast Algorithms for Convolutional Neural Networks ?

[jdnie]

G, B, A can be calculated by TOOM-COOK algorighm or Chinese remainder theorem, you can try to calculate F(4,3) F(6, 3)

[buttercutter]

but in your own screenshot, matrices G and B are derived from matrix A ?

[jdnie]

your G is my B, A is my G, and B is my A

[jdnie]

G transform g from 3 dims to 4 dims, B transform d from 4 dims to 4 dims, A transform Y from 2 dims to 4 dims, then use dot product replace convolution. It's the same as the FFT transform pics from spatial domain to frequency domain, and use dot product in frequency domain to replace the convolution in frequency domain. The G/A/B are the transform basises.

[buttercutter]

G transform g from 3 dims to 4 dims, B transform d from 4 dims to 4 dims, A transform Y from 2 dims to 4 dims

This above sentence is confusing.

your G is my B, A is my G, and B is my A

ok, but how you derive your G and your B from your A as shown in your screenshot ?

[jdnie]

In TOOM-COOK, you can make a set of equations and solve it.

[buttercutter]

@jdnie Could you advise about https://github.com/NervanaSystems/neon/issues/224#issuecomment-605602605 ?

[buttercutter]

@jdnie

What are β0 , β1 and β2 ?