Carlo Lucibello
Carlo Lucibello
i.e. extend #199 to all activations, now that Zygote implements the vmap adjoint. Let's do proper benchmarking when doing this. @AStupidBear
We should add what is called elsewhere `topk` or `maxk`, e.g. https://pytorch.org/docs/stable/generated/torch.topk.html In Base we have `partialsort` and `partialsortperm` doing something similar but limited to 1d arrays
While multiplication of generic matrices by Zeros is already handled, multiplication by Ones is not, while for Fill we only have right multiplication. As you can from the benchmarks below,...
It would be nice to have `Zeros(eltype, dims)` and `Ones(eltype, dims)`, in analogy to `zeros` and `ones` functions
Probably there is an issue already corresponding to this but I couldn't find it. We encountered this error in Fux https://github.com/FluxML/Flux.jl/issues/1935 but ultimately can be reduced to indexing a view...
On master, even the simplest reduction fails: ```julia julia> using CUDA, CUDA.CUSPARSE, Random, SparseArrays julia> x = sprand(3, 3, 0.5) |> CuSparseMatrixCSR 3×3 CuSparseMatrixCSR{Float64, Int32} with 4 stored entries: ⋅...
On CUDA#master: ```julia julia> using SparseArrays, CUDA, CUDA.CUSPARSE julia> CUDA.allowscalar(false) julia> A = CuSparseMatrixCSR(sprand(3, 3, 0.5)) 3×3 CuSparseMatrixCSR{Float64, Int32} with 5 stored entries: ⋅ 0.621676345112914 ⋅ 0.6600808319576792 0.05411482677405954 0.016950636692096288 ⋅...
Some of the features of the OGBDataset are downloaded as torch tensor stored in the ".pt" format. They are currently ignored at the moment, but we could load them using...
There is a lot of output when testing OGBDataset and SMSCollection, [see this CI run](https://github.com/JuliaML/MLDatasets.jl/runs/7156072504?check_suite_focus=true#step:5:611). Maybe we can suppress these outputs although I'm not sure why we don't see this...