Thu, 18 Apr 2024 10:51:10 +0300
commit before adding PET
#################################################################### # Predictive online PDPS for optical flow with known velocity field #################################################################### __precompile__() module AlgorithmPrimalOnly identifier = "pdps_known_primalonly" using Printf using AlgTools.Util import AlgTools.Iterate using ImageTools.Gradient using ..OpticalFlow: ImageSize, Image, pdflow! ######################### # Iterate initialisation ######################### function init_rest(x::Image) imdim=size(x) y = zeros(2, imdim...) Δx = copy(x) Δy = copy(y) x̄ = copy(x) return x, y, Δx, Δy, x̄ end function init_iterates(xinit::Image) return init_rest(copy(xinit)) end function init_iterates(dim::ImageSize) return init_rest(zeros(dim...)) end ############ # Algorithm ############ function solve( :: Type{DisplacementT}; dim :: ImageSize, iterate = AlgTools.simple_iterate, params::NamedTuple) where DisplacementT ################################ # Extract and set up parameters ################################ α, ρ = params.α, params.ρ R_K² = ∇₂_norm₂₂_est² γ = 1.0 Λ = params.Λ τ₀, σ₀ = params.τ₀, params.σ₀ τ = τ₀/γ @assert(1+γ*τ ≥ Λ) σ = σ₀*1/(τ*R_K²) println("Step length parameters: τ=$(τ), σ=$(σ)") ###################### # Initialise iterates ###################### x, y, Δx, Δy, x̄ = init_iterates(dim) init_data = (params.init == :data) #################### # Run the algorithm #################### v = iterate(params) do verbose :: Function, b :: Image, v_known :: DisplacementT, 🚫unused_b_next :: Image ################## # Prediction step ################## if init_data x .= b init_data = false end pdflow!(x, Δx, y, Δy, v_known, false) ############ # PDPS step ############ ∇₂ᵀ!(Δx, y) # primal step: @. x̄ = x # | save old x for over-relax @. x = (x-τ*(Δx-b))/(1+τ) # | prox @. x̄ = 2x - x̄ # over-relax ∇₂!(Δy, x̄) # dual step: y @. y = (y + σ*Δy)/(1 + σ*ρ/α) # | proj_norm₂₁ball!(y, α) # | prox ################################ # Give function value if needed ################################ v = verbose() do ∇₂!(Δy, x) value = norm₂²(b-x)/2 + params.α*γnorm₂₁(Δy, params.ρ) value, x, [NaN, NaN], nothing end v end return x, y, v end end # Module