You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
* Use astropy Tables for data everywhere
* Load npz file if present
* Improve cache coherence in NTSeason.simulate_background
Cuts time in Injector.create_dataset() from ~1500 ms to ~600 ms
* CircularGaussian: do not attempt to modify data in-place
* Blacken/isorten following #415
* Use Table for EffectiveAreaInjector
* Annotate factory functions
* Avoid modifying columns in place
* Use instance variables in NorthernTracksKDE
* Normalize KDE PDF to region of interest
* Speed up TableInjector
by a factor ~2 when drawing ~175 events from 11.5e6 MC events across 1000 sources
* Speed up StaticFloor
numpy broadcasting is much faster than a list comprehension
* Cache effective injection time for steady sources
There's no point in calculating this tens of thousands of times, since it's independent of the source for a steady time pdf
* Remove upper bound on n_s
* Squeeze the last bits of fluff out of StdMatrixKDEEnabledLLH.get_spatially_coincident_indices
Unexpected bottlenecks:
- Repeatedly slicing into astropy Tables (overhead from copying units)
- Temporary arrays in ra_dist and angular_distance
* Evaluate SoB for different gammas lazily
Keys appear as before, but values are only evaluated when accessed. Speeds up initialization when gamma is not being fitted.
* Use less accurate (but much faster) angular distance calculation in NorthernTracksKDE
* Annotate PDF factory functions
* HACK: allow StdMatrixKDEEnabledLLH to use CircularGaussian pdf
* Revert "Annotate PDF factory functions"
This reverts commit 9afa53e13556e783a73909fa2751ed98bad35f27.
* Normalize gaussian PDF
* WIP: King function PSF
* chore: annotate BaseInjector
* Allow Season to simulate background only in the source box
* Evaluate energy SoB lazily
* Add log of King PDF
* Add an option to disable energy S/B interpolation
The 2nd-order spline interpolation does not necessarily reflect the normalization of the underyling signal or background energy PDFs
* Use splines for King PDF parameters
* Speed up make_season_weight 10x
180 ms -> 11 ms for 22k sources
* Microoptimization: sparse matrix projection 25% faster
* Microoptimization: slice result of background pdf calculation instead of inputs
* mypy cleanup
* Guard annotation-only imports
* Cosmetic: raw string for latex
* Return excluded events from create_dataset in unblinded injectors
* Consider excluded events in flare likelihoods (but require ==0)
* Handle unbounded n_s in 1D llh scan
0 commit comments