hippke / planejade Goto Github PK

Features

• Test speed with partially cached log-likelhood (only re-calculate area in-transit)
• small planet approximation OK? For small planets, yes. For big: Sensitivity sufficient anyways. Shape features equivalents possible (both models can fit any curve perfectly, but with small-planet with distorted params?)
• save all points for cornerplots
• use numba prange to handle N batches in parallel. Show combined status (with ****). Migrate to procedural?
• konvergieren nicht nach iters, sondern nach volume (am ende wird's ganz klein)
• oder nach dlogz wie dynesty/ultranest
• nice progress like ultranest (jetzt doch)
• test recovery rate for SNR=10 for different periods
• Fast biweight filter
• dynesty sampling (fastest?)
• put tqdm to 100% after premature convergence
• GPU version (vectorized, vector size=pop), cuda_example.py
• CPU vectorized (vector size=pop/cores)
• offer with and without detrending
• optional cornerplots
• add seed=None to turboJADE
• Convert JADE class to function (to get rid of pathos dependency)
• shift t0 to first t0 in time series

Stats:

• FAP from SNR or Bayes
• logl erweitern um SNR, in-transit points, etc (siehe TLS); BIC vs. flat line model
• Biweight filter with memory of out-of-transit points (only re-calculate what is needed)
• Scatterplot: Bayes BIC vs. SNR for different planet radii (all else equal)
• BIC also for ecc if significant
• as in TLS: Phase-folded model for visualization purpose

Write paper:

• own numba JADE (both papers)

API:

• Allow to set bounds min=max --> Parameter is not free any more
• Get emcee fit after discovery?

Helpers:

• transit mask

Implemented - ready for testing:

• Modify T0 internally to be phase; should speed up convergence. Only output T0=t_min + phase(period)
• Add physical limits (a)
• numba
• Convergence cutoff after certain no_iters of no progress? Test n_it=2000 with cutof=100 for deep transits to check if converged
• Ensemble 16x250, 8x500, 4x1000 // good for all cases? Test low SNR and high SNR. Are high SNR converged with 250?
• SDE, Power Spectrum: SDE(period), get chi2 of each model eval...is periodogram as useful as in TLS? ==> not useful, because part of spectrum (some per) are not fully explored
• Pretty printing result stats, model vals
• (not implemented) Pretty printing status (like UltraNest)
• occult_hybrid with integrated loglike ==> not faster

---