Cuda support?

[kristopher-wood]

Does River have any Cuda support? It would be nice if I could fully utilize my GPU.

[chrsunwil]

Hi, I would love to bump this question. I assume that there would still be some speed up for these models (although I know that the largest speed-ups are general for neural networks).

[MaxHalford]

There are likely gains to be made by adding some parallelism in there. To confirm, do you notice if your CPU usage is close to a 100% or not?

[chrsunwil]

My CPU is only at about 23%, but I wasn't sure how much parallelism was built into the implementation? Also, to be fair, I'm training two models so I can parallelize my surrounding code a bit.

[MaxHalford]

There is no parallelism built in as of now :)

[chrsunwil]

Even for bagging ensembles? (Those are embarrassingly parallel, right?). Because my CPU is only 23%, that indicates there could be gains from parallelization, right?

As a side note, thanks for maintaining this package! I just discovered it a handful of weeks ago and it's really wonderful.

[MaxHalford]

Even for bagging ensembles? (Those are embarrassingly parallel, right?). Because my CPU is only 23%, that indicates there could be gains from parallelization, right?

There should be, in theory. But then in reality you have a communication cost: each observation needs to be sent to a worker for processing. Sometimes that results in too much overhead. There's no silver bullet, especially for small processing loads, as is the case with online machine learning!

As a side note, thanks for maintaining this package! I just discovered it a handful of weeks ago and it's really wonderful.

Cheers, we're doing what we can :)