[Feature]: How can I use GPU to accelerate training, like deep learning? about pysr HOT 5 CLOSED

Hi @guoaiBB,

The backend would have to some changes before it would support CUDA natively.

I think basically it would need to be able to store symbolic expressions as some CUDA arrays in a new Node{T}-like type that subtypes AbstractNode: https://github.com/SymbolicML/DynamicExpressions.jl/blob/master/src/Equation.jl. It might be a lot of work FYI.

But, if you're just using CUDA for array evaluation, then it shouldn't be so bad. It might already work (on the Julia side of things). Take a look at https://github.com/JuliaGPU/CUDA.jl.

Cheers,
Miles

from pysr.

Closing in favor of SymbolicML/DynamicExpressions.jl#31

from pysr.

from pysr.

Hi @guoaiBB,

If it's okay with you, could you post your questions separately? Generally I try to keep one topic per issue/discussion so it's easier for people to find if they search it.

For this question in particular, it's good if you could initially check the tuning advice page (https://astroautomata.com/PySR/tuning/) as well as the forums (https://github.com/MilesCranmer/PySR/discussions). If you can't find an answer there then feel free to post a new discussion for help.

(300k rows is too large for PySR; you generally only need 5k max)

Cheers,
Miles

from pysr.

from pysr.