Brendan O'Donoghue
Brendan O'Donoghue
Thanks, I'll investigate.
Thanks for running this! As the `acceleration_lookback` parameter increases the linear system it's solving gets worse conditioned, but the acceleration gets more effective because it uses more iterates to extrapolate...
Hi @nrontsis, yes please share the data if you can. You can email me if it's sensitive: [email protected] It's looking like the acceleration parameter is too high by default. I...
Thanks for investigating that @kalmarek. I haven't had much time to work on SCS lately but I'm hoping to do some investigation soon and this information is very helpful!
@kalmarek Thanks for running these, glad to see it's slightly more stable now. Some more stability improvements are in the pipeline.
Great! I think generally the guidance in the literature for the AA memory parameter is 5-10, but for some problems more memory can be beneficial. However too much memory accumulates...
Happy to hear that the GPU support is helping! Unfortunately at the moment SCS can only use one GPU, since it just ships the data over to a single GPU....
Is this something you're seeing for every problem or just one you're trying? Also make sure you're comparing the indirect CPU version (`use_indirect=True` in the python call).
Looks like you want higher accuracy here? If so you can reduce the `eps` quantity through cvxpy which will yield better solutions (but take more time to solve).
@BptGrm The errors about missing `#include ` are when you don't have cuda installed on your system, or if it is installed then the location of the headers is not...