-
Notifications
You must be signed in to change notification settings - Fork 0
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
use SimpleChains to train/run simple NNs #212
Comments
I didn't see a way to add comments to stuff on the project board without converting to issue first. I just wanted to add the context that SimpleChains doesn't have GPU support, which means it would excise the huge CUDA stack that gets compiled and loaded every time. Another alternative is waiting on FluxML/Flux.jl#2132. |
@anandijain do you know if Flux is the only package that has the CUDA dependency? |
im not sure, i can check before our next meeting. the way to do it would be to use my code to generate the registry graph, then enumerate all paths from FUSE.jl to CUDA and see if flux is always an intermediate. however, just looking at the SVG, I did see a GA package that had CUDA as a direct dep |
Ok, but function on_device(args)
if CUDA.functional() && args.use_cuda
@debug "Training on CUDA GPU"
CUDA.allowscalar(false)
return Flux.gpu
else
@debug "Training on CPU"
return Flux.cpu
end
end Probably the direct CUDA dependency can be removed by doing Meaning that if we removed the |
Yes, thats correct |
I am looking at the size of your NNs and it does seem like the right size for SimpleChains to be fast. I was talking to Chris (the dev of SimpleChains) and he cautioned that it works, but is a bit limited with features and documentation |
After discussing with @ChrisRackauckas we have decided this is not worth the effort at this time |
https://github.com/PumasAI/SimpleChains.jl
The text was updated successfully, but these errors were encountered: