Closed segasai closed 3 years ago
Changes Missing Coverage | Covered Lines | Changed/Added Lines | % | ||
---|---|---|---|---|---|
py/dynesty/bounding.py | 12 | 13 | 92.31% | ||
py/dynesty/dynesty.py | 0 | 3 | 0.0% | ||
<!-- | Total: | 12 | 16 | 75.0% | --> |
Totals | |
---|---|
Change from base Build 937043486: | -0.05% |
Covered Lines: | 3580 |
Relevant Lines: | 4637 |
It looks like now the patch recovers the high dim behaviour that was there before. It also removes the vol_dec parameters so I think it's ready to be reviewed/merged.
following #285 make the conditions stricter on hierarchical splits if the first split didn't work also get rid of no-op vol_check option.
This partially recovers the losses of #284 (blue curve) as the deviations start to now happen at ndim=17
I'm not sure we'll be able to fully get it back, as the previous code used pointvol for this. In my mind there were cases where pointvol estimates were good and it was preventing excessive splitting but in some cases it was preventing splitting because of incorrect pointvols.
I also change the sampler to rslice in high dims.