When rank is 70, it does not converge ······

[zhibin-daily]

Hi：
CuMF is very efficient. It is a amazing results.But I have three questions, the first is that RMSE does not converge when lambda is 0.05 and rank is 70,this is a very strange situation. The second is that should I process the matrix R into block form and Stored as CSR and CSC format before run the SU-ALS? The last is that can you send me the code of SU-ALS? I am very interested in this algorithm.

[wei-tan]

@zhibin-daily : we also noticed the issue of rank = 70, very recently. It is VERY strange and we are looking into it. AS for SU-ALS, depending on if you need model parallelism or data parallelism or both, you need to blocktize R from both CSC and CSRs. We use Python and numpy to do this before sending to GPUs. This part is not currently open sourced as we have not generalized it to deal with data of any size yet. But I can help if you have specific questions.

[Zialus]

Has there been any update on this issue?

[sh1ng]

There's a race condition in get_hermitianT10 and get_hermitianT100.
It's fixed in h2oai/h2o4gpu#729 Also python binding is added and we are going to maintain and improve the algorithm.