Closed nmheim closed 3 years ago
I definitely agree with the vector part. Not sure about matrices, but I guess your suggestion is the most logical. I just worried how to broadcast it. Can you think of a way how to write quadratic form x^T \Sigma^{-1} x with matrix3darraymatrix?
Sorry closed this with the pullrequest without checking in here first. We could overload multiplication, but I am not sure if this is what we want...
Currently the CMeanVarGaussian{ScalaVar} is returning a
(1,batchsize)
array for variance. we agreed with @smidl at some point that variance should always return a proper vector, so I will change this accordingly, such that we get(xlen,batchsize)
fromvar
. I assume we also want this to be the case for the covariance matrices, so we would end up with(xlen, xlen, batchsize)
. The memory overhead would be acceptable if we useDiagonal
s for the covariances. What do you all think? @vitskvara @pevnak