Sampler comparison about pandora HOT 6 CLOSED

UltraNest

Perhaps useful:

• ultranest.stepsampler.RegionSliceSampler(2*nparams) -- crashes, worked earlier

• ultranest.stepsampler.RegionSliceSampler(500) -- (8 epochs 2.5d LC) 10^8 evals in 5hrs, VERY GOOD result

• ultranest.stepsampler.RegionSliceSampler(50) -- 21min, nearly identical results

• ultranest.stepsampler.CubeSliceSampler(2*nparams) -- 10min, slightly wrong

• ultranest.stepsampler.CubeSliceSampler(50) -- slightly wrong

• ultranest.stepsampler.CubeSliceSampler(500) - same slightly wrong

Not useful:

• ultranest.stepsampler.RegionBallSliceSampler -- very wrong
• ultranest.stepsampler.BallSliceSampler(nsteps=nsteps) -- slow, wrong
• ultranest.stepsampler.CubeMHSampler(nsteps=nsteps) -- Slow, very wrong
• ultranest.stepsampler.RegionMHSampler(nsteps=nsteps) -- Very fast, very wrong
• ultranest.ReactiveNestedSampler -- very slow (days)

Speed UltraNest MPI

• 1 Thread: 5000/sec
• 4 Threads: 8700/sec
• 8 Threads: 8000/sec

from pandora.

dynesty
Testing with 4 epochs each 2.5 days of LC data

• Crashes when printing results

• Plotting results OK

• DynamicNestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='single') -- 6min, good results (broad, most correct)

• DynamicNestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='multi') -- 12min, results worse (multimodal)

• DynamicNestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='multi', nlive=4000) -- 8min, wrong peak

• DynamicNestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='multi') -- 1hr 38, nah an der wahren Lösung, bestes Ergebnis bislang

• DynamicNestedSampler(
  log_likelihood,
  prior_transform,
  ndim=ndim,
  bound='multi',
  sample='hslice',
  slices=10) ==> 10hrs, 2.5e8 evals, 6,950 evals/sec! Result very bad (tight, wrong peaks)

Current raw throughput (prange Pandora segments, 10 epochs, 2.5d, Kepler cadence) is 8000 models /sec

• DynamicNestedSampler(
  log_likelihood,
  prior_transform,
  ndim=ndim,
  bound='none',
  sample='hslice',
  slices=10,
  gradient=gradient) ==> 7h 1.7e8 evals, alles total falsch

• NestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='single') -- 3min, broad good results

• NestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='multi') -- 4min, broad good results

• NestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='multi', nlive=1000) -- 6min, broad, worse results

• NestedSampler(log_likelihood, prior_transform, ndim=ndim, bound='multi', nlive=4000) -- 22min, worse

Speed Dynesty:

• 1 Thread: 5000/sec
• 8 Threads: 17,000/sec (after burn-in)

sample="rslice" -- gets stuck, crashes
sample="rstagger" --
sample="hslice" -- gets stuck, crashes

from pandora.

Nestle

• cancelled after 18 hrs

from pandora.

Zeus
No useful prior transformation concept (?)

from pandora.

UltraNest
sampler.stepsampler = ultranest.stepsampler.RegionSliceSampler(
nsteps=4000,
adaptive_nsteps='move-distance'
)
result = sampler.run(min_num_live_points=4000)
fig_retrieval_v4.pdf

Results:
logZ = -1802.250
40h runtime, 414m evals
near-perfect lightcurve
per_moon: wrong peak

from pandora.

Test dyPolyChord

from pandora.