This package allows algorithms developed with AbstractImageReconstruction.jl to be exeucted partially or fully in different (and potentially distributed) Julia processes using Dagger.jl
DaggerImageReconstruction.jl is a Julia package that utilizes Dagger.jl to distribute image reconstruction algorithms across multiple Julia processes, either partially or fully. This package is designed to extend the generic interface of AbstractImageReconstruction.jl.
You can install the package using Julia's package manager. Open the Julia REPL and run:
using Pkg
Pkg.add("DaggerImageReconstruction")To use DaggerImageReconstruction.jl, one first needs to add a new Julia process using Distributed:
# Add new process (on remote server)
using Distributed
addprocs(["gpuServer"])Afterwards one can load `DaggerImageReconstruction´ and the packages implementing specific image reconstruction algorithms, such as [MPIReco.jl]. Similar to AbstractImageReconstruction.jl, this package does not offer concrete reconstruction algorithms:
using DaggerImageReconstruction
@everywhere using AbstractImageReconstruction, ... # Load Reco packages
plan = loadDaggerPlan("plan.toml", ...; worker = 2)
# Transparently configure algorithm in the remote process
dict = Dict{Symbol, Any}()
dict[:iterations] = 20
dict[:file] = () -> File("path/on/remote")
setAll!(plan, dict)
data = ... # read local data
# Transfer data to the remote, reconstruct there and return the result
algo = build(plan)
image = reconstruct(algo, raw)Algorithms can be parameterized and constructed as if they were local RecoPlans. The usage example moves the whole image reconstruction to another process. DaggerImageReconstruction also offers an AbstractUtilityReconstructionParameters to move individual process steps to another process.
Contributions are welcome! If you would like to contribute to DaggerImageReconstruction.jl, please fork the repository and create a pull request.
This package is licensed under the MIT License. See the LICENSE file for more information.
This package is built on top of Dagger.jl, which provides the framefork for the distributed computations.