search

wesselb / gpar

Implementation of the Gaussian Process Autoregressive Regression Model
MIT License
59 stars 10 forks source link

GPU support #2

Open v-a-s-a opened 4 years ago

v-a-s-a commented 4 years ago

I'm fitting a mid-sized time series model: two observations at 8k time points. There are some optimizations already available in GPAR, e.g. this can be fit with ~6k inducing points with some reasonable assumptions.

In the interest of performance, is it possible /advisable to try and enable GPU on the wrapped torch or tensorflow backends? The stheno, lab and plum-dispatch dependencies don't mention GPU support, but the flexibility they provide is tantalizing.

wesselb commented 4 years ago

@v-a-s-a I think GPU support should actually not be too difficult, and almost work out of the box. The main issue is, e.g. in Torch, that new tensors created are not automatically moved to the GPU, but this can of course be fixed. I will look into this over the next week!

v-a-s-a commented 4 years ago

That sounds fantastic! Let me know if there is any testing/benchmarking to be done. I'm curious if there are any appreciable gains to be had.

wesselb commented 4 years ago

@v-a-s-a It turns out to be a busy time right now; I might need a few weeks to sort this out. I'll keep you up to date.

v-a-s-a commented 4 years ago

Sounds great @wesselb. I was poking around and wondering how easy this would be to setup. The GPU support isn't strictly necessary for me needs here. Feel free to close/open as you see fit.