You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
This repository has been archived by the owner on Sep 11, 2023. It is now read-only.
Apparently, there is a problem with some dependencies that cause the eigendecomposition of sparse matrices to fail. In our debug-session, @marscher figured out that at least with newer versions of scipy / openblas, unit tests with sparse matrices fail. Eigendecompositions cannot be computed or produce random results. The issue will be reported upstream.
A conda environment yaml file that reproduces the issue is attached here.
However, the issue could be reproduced with any clean installation of pyemma.
The problem can be avoided by using dense matrices (default behavior).
The text was updated successfully, but these errors were encountered:
We pinned down the problem to a specific version of OpenBLAS, namely 3.8 from conda-forge. It seems like build number 7 is broken, while number 10 has been tested to work with PyEMMA.
>>> conda install blas
The following NEW packages will be INSTALLED:
blas conda-forge/linux-64::blas-2.10-mkl
liblapacke conda-forge/linux-64::liblapacke-3.8.0-10_mkl
The following packages will be UPDATED:
libblas 3.8.0-7_openblas --> 3.8.0-10_mkl
libcblas 3.8.0-7_openblas --> 3.8.0-10_mkl
liblapack 3.8.0-7_openblas --> 3.8.0-10_mkl
openblas 0.3.5 had issues with SkyLake X cpus. dgemm was broken there.
I'll move 0.3.5 openblas to broken
This means, that we will not get any more broken installations, which is very good. But I have no clue how to deal with the problem, that some installations are already broken. We can not reach people properly and in the meantime they will have bogus results.
marscher
changed the title
sparse matrices may be broken
Solutions of eigenvalue problems yield bogus results
Jul 2, 2019
marscher
changed the title
Solutions of eigenvalue problems yield bogus results
Solving of eigenvalue problems yield bogus results with OpenBLAS
Jul 2, 2019
marscher
changed the title
Solving of eigenvalue problems yield bogus results with OpenBLAS
Solving of eigenvalue problems yield bogus results with OpenBLAS-0.3.5 on SkyLake X CPUs
Jul 15, 2019
Apparently, there is a problem with some dependencies that cause the eigendecomposition of sparse matrices to fail. In our debug-session, @marscher figured out that at least with newer versions of scipy / openblas, unit tests with sparse matrices fail. Eigendecompositions cannot be computed or produce random results. The issue will be reported upstream.
A conda environment yaml file that reproduces the issue is attached here.
However, the issue could be reproduced with any clean installation of pyemma.
The problem can be avoided by using dense matrices (default behavior).
The text was updated successfully, but these errors were encountered: