Skip to content

Allocations in ADMM #107

New issue
New issue
Open
#108
Open
Allocations in ADMM#107
#108
@JakobAsslaender

Description

@JakobAsslaender
JakobAsslaender
opened on Sep 26, 2025

The calculations of all convergenance criteria are allocating:

RegularizedLeastSquares.jl/src/ADMM.jl

Lines 247 to 256 in cff4e12

state.rᵏ[i] = norm(solver.regTrafo[i] * state.x - state.z[i]) # primal residual (x-z)
state.sᵏ[i] = norm(state.ρ[i] * adjoint(solver.regTrafo[i]) * (state.z[i] .- state.zᵒˡᵈ[i])) # dual residual (concerning f(x))
state.ɛᵖʳⁱ[i] = max(norm(solver.regTrafo[i] * state.x), norm(state.z[i]))
state.ɛᵈᵘᵃ[i] = norm(state.ρ[i] * adjoint(solver.regTrafo[i]) * state.u[i])
Δᵒˡᵈ = state.Δ[i]
state.Δ[i] = norm(state.x .- state.xᵒˡᵈ ) +
norm(state.z[i] .- state.zᵒˡᵈ[i]) +
norm(state.u[i] .- state.uᵒˡᵈ[i])

I have noticed that ADMM reconstructions tend to have a memory usage that grows with the iterations, and I presume it is due to these lines I tested it in one example case, and commenting out these lines fixes the issue. Of course, this prevents adaptive rho and convergance tracking. Any idea how to fix this @nHackel?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions