add ChainRulesCore rules

[mileslucas]

This PR adds analytical gradients using ChainRulesCore.jl

[codecov]

Codecov Report

Merging #3 (eb7de41) into main (1399529) will not change coverage.
The diff coverage is n/a.

❗ Current head eb7de41 differs from pull request most recent head d46340a. Consider uploading reports for the commit d46340a to get more accurate results

@@           Coverage Diff           @@
##             main       #3   +/-   ##
=======================================
  Coverage   98.80%   98.80%           
=======================================
  Files           6        6           
  Lines          84       84           
=======================================
  Hits           83       83           
  Misses          1        1           

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 1399529...d46340a. Read the comment docs.

[mileslucas]

I don't understand why the chain rule tests are failing. Let's look at the isotropic Gaussian PSF as an example

Here is the definition of the gradient

PSFModels.jl/src/gaussian.jl

Lines 67 to 77 in 1b026c5

	# isotropic
	function fgrad(g::Gaussian, point::AbstractVector)
	f = g(point)
	xdiff = first(point) - first(g.pos)
	ydiff = last(point) - last(g.pos)
	dfdpos = -2 * GAUSS_PRE * f / g.fwhm^2 .* SA[xdiff, ydiff]
	dfdfwhm = -2 * GAUSS_PRE * f * (xdiff^2 + ydiff^2) / g.fwhm^3
	dfdamp = f / g.amp
	return f, dfdpos, dfdfwhm, dfdamp
	end

which I wrote out by hand and can be verified with this derivation http://umdberg.pbworks.com/w/page/88516931/Example%3A%20Gradient%20of%20a%20Gaussian

here are the chain rules

PSFModels.jl/src/gaussian.jl

Lines 94 to 111 in 1b026c5

	function frule((Δpsf, Δp), g::Gaussian, point::AbstractVector)
	f, dfdpos, dfdfwhm, dfda = fgrad(g, point)
	Δf = dot(dfdpos, Δpsf.pos) + dot(dfdfwhm, Δpsf.fwhm) + dfda * Δpsf.amp
	Δf -= dot(dfdpos, Δp)
	return f, Δf
	end
	function rrule(g::G, point::AbstractVector) where {G<:Gaussian}
	f, dfdpos, dfdfwhm, dfda = fgrad(g, point)
	function Gaussian_pullback(Δf)
	∂pos = dfdpos .* Δf
	∂fwhm = dfdfwhm .* Δf
	∂g = Tangent{G}(pos=∂pos, fwhm=∂fwhm, amp=dfda * Δf, indices=ZeroTangent())
	∂pos = dfdpos .* -Δf
	return ∂g, ∂pos
	end
	return f, Gaussian_pullback
	end

using them works as intended-

using ChainRulescore, PSFModels
psf = PSFModels.Gaussian(fwhm=10)
point = [1, 2]
f, pullback = rrule(psf, point)
Δpsf, Δpoint = pullback(1.0)
f2, Δf = frule((Δpsf, Δpoint), psf, point)

# output
(0.8705505632961241, 0.7817442466933209)

but using test_frule and test_rrule consistently fails

PSFModels.jl/test/runtests.jl

Lines 99 to 111 in 1b026c5

	@testset "gradients" begin
	# have to make sure PSFs are all floating point so tangents don't have type issues
	psf_iso = Gaussian(fwhm=10.0, pos=zeros(2))
	psf_tang = Tangent{Gaussian}(fwhm=rand(rng), pos=rand(rng, 2), amp=rand(rng), indices=ZeroTangent())
	point = Float64[1, 2]
	test_frule(psf_iso ⊢ psf_tang, point)
	test_rrule(psf_iso ⊢ psf_tang, point)
	psf_diag = Gaussian(fwhm=Float64[10, 8], pos=zeros(2))
	psf_tang = Tangent{Gaussian}(fwhm=rand(rng, 2), pos=rand(rng, 2), amp=rand(rng), indices=ZeroTangent())
	test_frule(psf_diag ⊢ psf_tang, point)
	test_rrule(psf_diag ⊢ psf_tang, point)
	end

• frule failure
• rrule failure