This is an issue with the underlying divide and conquer algorithm used by
the LAPACK routine we are calling. It has been reported at Numpy's list as
well.
I think we should have an option for choosing the algorithm in svdfact,
such that you can easily switch to a QR based solver when the DnC solver
fails. Please open an issue with this. For at short term solution, you can
get your values and vectors by directly calling LAPACK.gesvd!('S','S',A)
where A is your matrix.
2015-01-30 6:02 GMT-05:00 Steven Sagaert <[email protected]>:
> when doing an SVD of a large matrix I get
> ERROR: LAPACKException(1)
> in gesdd! at linalg/lapack.jl:1046
> in svdfact! at linalg/factorization.jl:660
> in svdfact at linalg/factorization.jl:664
>
> It's definitely something related to the data because it works on
> different matrices & the code has been working for months. Any idea what
> that error is about?
>
> I'm working on Ubuntu 14.04, julia version 0.3.5
>
