added anscombe and tests and changed initial scaling #20
Conversation
| CallCounter = 0; | ||
| sqrt_meas_b = 0; | ||
| function call_brancher(μ, meas, storage=copy(µ)) # uses a closure to deal with possibly normalized measurement data. | ||
| if CallCounter == 0 |
There was a problem hiding this comment.
Call Counter can be removed since Optim.jl takes track of how often the functions are called (fg!)
There was a problem hiding this comment.
The role of the CallCounter is to ensure that the sqrt(measured + b) is calculated only once. However this is hacky and dangerous! I would prefer a different mechanism, but this means that the core of DeonvOptim needs a larger change.
| @@ -1,7 +1,8 @@ | |||
| export Poisson, poisson_aux | |||
| export Gauss, gauss_aux | |||
| export ScaledGauss, scaled_gauss_aux | |||
|
|
|||
| export Anscombe, anscombe_aux | |||
| using Tullio | |||
| end | ||
| storage .= abs2.(sqrt_meas_b .- sqrt.(μ .+ b)) | ||
| return sum(storage) | ||
| #@tullio s = @inbounds abs2(sqrt_meas_b[i] - sqrt(μ[i] + b)) |
There was a problem hiding this comment.
? you mean only the commented tullio line, I guess? Yes, but I prefer to keep it since this is what the future should eventually do, if tullio runs properly with Cuda.
| function ChainRulesCore.rrule(::typeof(anscombe_aux), μ, meas, storage=copy(μ); b=1, sqrt_meas_b=0) | ||
| Y = anscombe_aux(μ, meas, storage; b=b, sqrt_meas_b=sqrt_meas_b) | ||
| function anscombe_aux_pullback(xbar) | ||
| # @show xbar |
There was a problem hiding this comment.
Do you mean the comment with the show. Yes. remove
| Y = anscombe_aux(μ, meas, storage; b=b, sqrt_meas_b=sqrt_meas_b) | ||
| function anscombe_aux_pullback(xbar) | ||
| # @show xbar | ||
| # @tullio storage[i] = @inbounds (one(eltype(μ)) - sqrt_meas_b[i] / sqrt(μ[i] + b)) # xbar[i] |
There was a problem hiding this comment.
show yes, tullio I prefer to keep.
this adds the anscombe transform. I am not entirely happy with the way that the persisitent memory is dealt with by trapping the first call to the function within the deconv iterations...