vedantchandra / wdtools Goto Github PK

Rather than un-pickling an old random forest, check the versions and delete and re-train the RF ensemble if sklearn updates.

Hello, I've noticed that there is a systematic shift of ~20 angstroms in the H-beta peak from center for all of the fits that I've done. Just as a test, we ran the same spectrum through wdmodel and this shift was not present there. I've attached a sample fit for reference. Not sure if this is an easily fixable issue or if others have experienced this as well.
sample_fit.pdf

nelder-mead is quite slow and inefficient. implement nonlinear least squares like LM to speed things up, and maybe even get a covariance matrix out of it. switch fitting to lmfit regardless to handle bounds. also, implement basinhopping or some similar global optimizer to replace the hot-cold rule of thumb.

We should incorporate a simple way to convert the baseline standard from air to vacuum wavelengths to accommodate different spectroscopic facilities.

After 'python setup.py install' on a m2 chip MacOS in a python3.8.18 environment, "import wdtools" will show this "failed to import GFP. only gaia subroutines will work" message, causing error of "AttributeError: module 'wdtools' has no attribute 'GFP'" after calling "gfp = wdtools.GFP(resolution = 2, specclass = 'DA')".
Any solution to this matter? Thanks~

Hello, I was using the wdtools package and struggled for a an hour or so to realise the issue. When you insert a spectrum to the fit_spectrum routine and the spectrum does not have halpha in its wl range, the routine crashes.
I realised later that rv_line and lines both need to be changed as an input, but I think that it would be useful for the next person to either be told that by the code when wl does not cover the range of e.g. 6000-7000A at all, or to script the routine to choose a balmer line that is in the wl range input and ignore halpha from 'fit_spectrum(lines='.
From the first few uses, the library looks really efficient and useful, producing a Teff and logg that I was expecting for some spectra. thanks!

this may be a tough one, but eventually, we should fit the entire spectrum and not just the Balmer lines. The level of sophistication here can be quite high, as in Narayan (2019).

combined spectro-photometry with a combined likelihood.

particularly to check dependencies and versions etc. could run a few simple fits and compare labels too to ensure the science side checks out.

Using an improved fitting technique it may be possible to extend the RF regression to H-epsilon and H8 as well.

simple bug where emcee doesn't work properly if the parameters are near their bounds (when the prior goes to zero). need to fix this on the forward-modelling side.