Fabian Pedregosa

I'm running the examples code with and without this patch, and the 12 CPUs I have are all constantly at 100% , and don't see any speed improvements after using...

I'm going to run a couple more benchmarks, but if the performance is the same I would be more inclined to let numba decide which parts to parallelize, as he'll...

and the parallelization I observe definitely comes from numba, as I can get it down to use just one CPU with the environment variable NUMBA_NUM_THREADS=1

OK I think I found why I was seeing all CPUs being used. It's because the bottleneck of the algorithm is not in the algorithm itself but in computing the...

thanks @cruyffturn ! Would you be willing to send a pull request fixing this issue?

that would indeed be cool. Is there any published work on this? Perhaps @gideonite knows

potentially relevant: https://arxiv.org/abs/2009.14794 On Wed, Jan 20, 2021 at 1:45 AM Geoffrey Negiar wrote: > See project proposal here > > . > > — > You are receiving this...

yes please :-)

that would indeed be awesome!

yes, you can exclude things that we don't want to test, like the data loading.