The GPU version works but slower than the paper reported.

[vicdxxx]

We wrote the below code working under Julia 12.1. Hope it can help further studies since the default code snippets cannot work directly. It takes 30 min for 10k samples on our PC-level GPU.

import Pkg
Pkg.activate(".")
Pkg.instantiate()

using Pkg; Pkg.add("NPZ")
using NPZ

using Dates
timestamp = Dates.format(now(), "yyyy_mm_dd_HH_MM_SS")

using KSVD

using Random, StatsBase, SparseArrays, LinearAlgebra
using KSVD.OhMyThreads, CUDA

include("ext/KSVDCudaExt.jl")
using .KSVDCudaExt

nsamples = 10_000
nnzpercol = 20
m = 4096
k = 20
emb_size = 2381
T = Float32
D = rand(Float32, emb_size, m)
X = stack(
(SparseVector(m,
sample(1:m, nnzpercol; replace=false),
rand(T, nnzpercol))
for _ in 1:nsamples);
dims=2)
Y = D*X + T(0.05)randn(T, size(DX))

println("use synthetic features")
Y_filename = "Y.npy"
println(Y_filename)

X_filename = "X.npy"
D_filename = "D.npy"
println(X_filename)
println(D_filename)

println("Y size: ", size(Y))

sparse_coding_method = KSVD.CUDAAcceleratedMatchingPursuit(max_nnz=k)

ksvd_update_method = KSVD.BatchedParallelKSVD{
false,
Float32,
KSVD.OhMyThreads.DynamicScheduler,
KSVD.CUDAAcceleratedArnoldiSVDSolver{Float32}
}(; shuffle_indices=true, batch_size_per_thread=1)

println("ksvd is starting")
(; D, X) = ksvd(Y, m;
ksvd_update_method,
sparse_coding_method,
maxiters=100,
abstol=1e-6,
reltol=1e-6,
show_trace=true)

err = mean(norm.(eachcol(Y - D * X)))
println("Mean reconstruction error = $err")

npzwrite(Y_filename, Y)_

npzwrite(X_filename, X)
npzwrite(D_filename, D)

println("Done!")