Closed mbrookhart closed 5 years ago
Yeah, after the blog post, people had matrices of parameters so we added a reshape in there to handle the different sizes. It looks like CuArrays can't handle reshaping a vector to a vector, and so that is the upstream issue you're pointing to.
This should be fixed on the master branch of CuArrays.jl. Check that out and see if there's any issues.
Thanks! I'll check it tonight.
Thanks! I had to update CUDAdrv and CUDAnative as well, but that got me past the original error. Now I'm seeing a different error in the Adam update of flux, but I wonder if it's just a version mismatch somewhere.
DiffEq had a breaking v6.0 release and we're looking at this in the next few days to make sure it gets updated.
First off, awesome paper, thank you!
When I try to run the NeuralODE example on CPU, it works great. However, when I try to switch it over to the GPU using the approach described in the paper i.e.,
using CuArrays
andx->neural_ode(gpu(dudt),gpu(x),tspan,BS3(),saveat=0.1)
, I see:This matches the error seen here: https://github.com/JuliaGPU/CuArrays.jl/issues/161
I'm using Julia 1.1, latest releases of all packages, CUDA 10 on a GTX 1070Ti on Ubuntu 18.04
Thanks!