fbartolic / caustic Goto Github PK

Matern32 is better than nothing but it doens't fully capture the autocorrelation in the residuals.

This is very important, need to implement finite sources with flexible limb darkening profiles first.

Finite source effects for single lens effects involve solving elliptical integrals which is costly. Current modeling codes use pre-computed tables for these elliptic integrals and interpolate a 1D function. If this approximation is fine it should be straightforward to implement in PyMC3, although instead of calling a table, I should fit a GP or a polynomial instead. Relevant references are here and here.

I don't wanna go to far with this. It should be able to handle OGLE/MOA/KMT/LCO data, it should support addition of various datasets. Need to rely on Astropy as much as possible.

Currently, the error bars given by surveys are rescaled by single free parameter, this is certainly wrong, need to figure out something less wrong.

Adding an additional constant variance term is probably good, figuring out how to capture flux dependant change in variance is better. Going hierarchical for error bar rescaling factors is probably over the top.

Need to add support for the dynesty dynamic nested sampling sampler to compare HMC with nested sampling on multi-modal parallax models.

I'm creating this issue because your repo contains a conf.py file with

nbsphinx_prompt_width = 0

... presumably in an attempt to remove the prompts from code cells.

I just want to give you a heads-up that this will stop working once spatialaudio/nbsphinx#439 is merged.

For more reliable alternatives, see https://nbsphinx.readthedocs.io/en/latest/custom-css.html.

Implement tests to check that the sampler is behaving correctly. Ideally, this should be done using Simulation Based Calibration.

Need to include annual parallax, it should be included by default in all models. Parametrization and priors are going to matter a lot, there are also potential problems with multimodality.

Plotting functions take up way too much memory and are often as slow as the sampling itself! There has be something I'm doing wrong.

This will take some time, need to sort out code structure and solve modeling problems first.

Following Rodrigo's comments. Should be the default for each model

Add support for observations in multiple bands for a given event.

Most models defined in Models.py are nested, the natural way to incorporate this fact in the code is to use class inheritance. For some reason it didn't work the last time I tried it with PyMC3 so at the moment I just duplicate code.

This issue seems relevant.

Implement least bad sigma clipping, or Gaussian mixture model.

There's almost no documentation at the moment, obviously this needs to change.