Henri Dehaybe
Henri Dehaybe
@tyler-ingebrand, you can test that hypothesis by copying the hyperparameters in the experiment I found here : https://github.com/zhihanyang2022/pytorch-sac If identical HPs do not solve this then a problem in the...
> Would it make sense to make something like QNetworkWithTarget
Just a heads up, I will resume committing to this PR now that Trajectories has been implemented. Several changes are needed to reach a mergeable state.
Indeed, there is a sign issue. I get a negative determinant for a PSD matrix: ``` LinearAlgebra.det(m::CuMatrix) = prod(diag(CUSOLVER.getrf!(m)[1])) #Shameless type piracy A = rand(10,10) S = A*A' S_d =...
> > PS. This is not "type piracy" because the `CuMatOrAdj` type is specific to this package. > > I guess he meant that this is type piracy in _his_...
All of the implementations proposed here are much slower than a CPU equivalent I fear. I made this benchmark where I compute the logdet without needing the LU decomposition (I...
I guess the overhead of `diag` allocating on the GPU is insignificant when one gets the GPU speed up for the decomposition. My use case is probably too specific to...
Is there any news on this issue ? I cannot exploit data I generated with an experiment due to this error. It is on a cluster so I have no...
Yeah that’s perhaps what I’ll have to do. Thought it is unideal because the “dataset” is a collection of trained neural nets that I will need to load to do...