Question: multiple warping?/ optimization of deformation?

[mdhe1248]

Hi Tim and Cody,
I am wondering how I can perform "two-step" registration. I first want to register images with a coarse grid size. Then, I want to use a high grid size for the second registartion. My strategy was to re-register warped images. Tim mentioned that we don't have to register images twice. There are something about optimization and initial guess. If I understood correctly, the deformations from the first registration can be used as an initial guess in the second registration.

After registration, I examined mon.

julia> mon[1]
Dict{Symbol,Any} with 1 entry:
  :u => 3x3x2 SharedArray{FixedSizeArrays.Vec{3,Float64},3}:…

I also obtained ϕs using griddeformations

u = load(fileout, "u")
ϕs = griddeformations(u, knots)

ϕs are deformations. How can I use the ϕs as an initial guess?

[timholy]

It's possible you'll need to make a couple of code tweaks in various places, but the hard work is already done. See, for example, https://github.com/HolyLab/BlockRegistration/blob/a328e76cf2a313caf68231ec387a1838f54b2a17/src/RegisterOptimize.jl#L509-L515.

[mdhe1248]

The line says,
ϕs, penalty = optimize!(ϕs, ϕs_old, dp, λt, mmis; kwargs...)
I guess I should decide λt by myself. Some experiment may be necessary. ϕs may be an empty array. ϕs_old can be obtained using u and griddeformations function. How could I obtain mmis and dp when my algorithm is RegisterWorkerApertures?

I obtained mismatch by adding :mismatch argument in monitor:

mon = monitor(algorithm, (), Dict{Symbol,Any}(:u=>ArrayDecl(Array{Vec{3,Float64},3}, gridsize), :mismatch => 0))

After registration, I got mismatch.

julia> mon[1]
Dict{Symbol,Any} with 2 entries:
  :u        => 3x3x2 SharedArray{FixedSizeArrays.Vec{3,Float64},3}:…
  :mismatch => 770315.6484996271

Can I use :mismatch as mmis in optimize! function?

Next, I tried to get dp:

mon = monitor(algorithm, (), Dict{Symbol,Any}(:u=>0, :datapenalty=>0))

After driver, the results do not seem promising:

julia> mon[1]
Dict{Symbol,Any} with 2 entries:
  :u           => 3x3x2 Array{FixedSizeArrays.Vec{3,Float64},3}:…
  :datapenalty => 0

I don't get datapenaly. It seems that datapenalty can be obtained when λ has a range.
https://github.com/HolyLab/BlockRegistrationScheduler/blob/master/src/RegisterWorkerApertures.jl#L180-L190
I tried a range of λ and got this error:

ERROR: On worker 2:
Too few data points for sigmoidal fit
 in fit_sigmoid at /home/donghoon/.julia/v0.4/BlockRegistration/src/RegisterOptimize.jl:994
 in auto_λ at /home/donghoon/.julia/v0.4/BlockRegistration/src/RegisterOptimize.jl:774
 in worker at /home/donghoon/.julia/v0.4/BlockRegistrationScheduler/src/RegisterWorkerApertures.jl:183
 in worker at /home/donghoon/.julia/v0.4/BlockRegistrationScheduler/src/RegisterWorkerShell.jl:148
 in anonymous at multi.jl:913
 in run_work_thunk at multi.jl:651
 [inlined code] from multi.jl:913
 in anonymous at task.jl:63
 in remotecall_fetch at multi.jl:737
 [inlined code] from /home/donghoon/.julia/v0.4/BlockRegistrationScheduler/src/RegisterDriver.jl:85
 in anonymous at task.jl:447

...and 3 other exceptions.

Then again, how could I obtain mmis and dp when my algorithm is RegisterWorkerApertures? Is there easier ways, where I don't need these variables?