Trouble plotting results of a dynesty run

[mas0nr]

Hello Phoebe community,

I've encountered some trouble making use of the dynesty backend and would appreciate any guidance on the issue.

I'm attempting to use dynesty to sample a light curve that has been phase-folded and then binned to 100 points. I've succesfully completed 2 seperate runs of dynesty on an HPC and retreived both as entire bundles for inspection on my local machine. When attempting to view the corner plot viab[solution_name].plot(style = 'corner') I am served with

ValueError                                Traceback (most recent call last)
Cell In[8], line 1
----> 1 b['first_dynest_run'].plot( style='corner')

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/parameters/parameters.py:5901, in ParameterSet.plot(self, twig, **kwargs)
   5897 if len(plot_kwargss) > 1:
   5898     # TODO: could we just return multiple figure instances?
   5899     raise ValueError("corner plots not supported with other axes.  Adjust the filter to include only a single distribution (including from compute or solver contexts), or to exclude all distributions.")
-> 5901 mplfig = plot_kwargs['dc'].plot(show=show, **plot_kwargs)
   5903 if 'failed_samples' in plot_kwargs.keys():
   5904     mplfig = _plot_failed_samples(mplfig, plot_kwargs.get('failed_samples', {}))

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/dependencies/distl/distl.py:7806, in MVSamples.plot(self, *args, **kwargs)
   7803 if dimension is not None:
   7804     kwargs.setdefault('bins', self.bins[dimension])
-> 7806 return super(MVSamples, self).plot(*args, **kwargs)

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/dependencies/distl/distl.py:3427, in BaseMultivariateDistribution.plot(self, **kwargs)
   3425     return self.slice(dimension).plot(**kwargs)
   3426 else:
-> 3427     return self.plot_sample(**kwargs)

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/dependencies/distl/distl.py:3398, in BaseMultivariateDistribution.plot_sample(self, size, **kwargs)
   3395     kwargs.setdefault('quantiles', (_norm.cdf(-1), _norm.cdf(1)))
...
    160             )
    162 else:
    163     force_range = True

ValueError: It looks like the parameter(s) in column(s) 0, 1, 2, 3, 4, 5 have no dynamic range. Please provide a `range` argument.

and a grid of empty plots. I'll provide my initial values, prior distributions, and solver options for reference

b.add_distribution(
    {
        'incl@binary':pb.gaussian_around(0.5),
        'q':pb.gaussian_around(0.3),
        'sma@binary':pb.gaussian_around(0.3),
        'teffratio@component':pb.gaussian_around(1),
        'teff@primary@component':pb.gaussian_around(2),
        'secondary@component@requiv':pb.gaussian_around(0.8)
    },
    distribution = 'june_10_dynesty',
    overwrite_individual=True
)
---
june_10_params = (
   10.27126516356737,   # incl
   0.8138771237102742, # q
   4.373635286772862, # sma
   0.7,                                 # teffratio
 8291,                               # teff primary
 1.74,                                 # requiv primary
 1.                                      # requiv 2
)
---
b.add_solver('sampler.dynesty',overwrite=True,
              solver='dynest_solver',  
              compute='phoebe01', 
              priors='june_10_dynesty',
              nlive='500,
              maxiter='10000',
              maxcall='20000',
              progress_every_niters = 100,
              
)

It may be worth noting that I have had to switch my modeling approach to this system from a semi-detached ellipsoidal variable to a detached system with a low inclination, as the former resulted in un-physical masses for both components. The agreement of the model and the observations suffered accordingly but I wanted to bring this system through the entire inverse problem regardless as I am new to the software and could use the experience. Below are plots of the binned light curve (black) and synthetic light curve (blue) to illustrate the fit before and after the second run:

Before

After

The solution recorded only 2446 iterations before it exceeded the maxcall limit so this will have to be increased in future runs, but it still seems strange to me that the posterior distributions can't be analyzed despite the successful completion of the solver.

The way I see it I might have to abandon hope for this system and move onto another that is clearly eclipsing but first I'd like to ask if anyone could shed some light on my next steps in dealing with this.

Thank you in advance for your time!

[kecnry]

Thanks for reaching out! I'm not sure at a quick glance exactly what is happening, but here are a few thoughts to consider:

• Dynesty isn't nearly as widely used or tested as the emcee sampler, so you may run into a few more issues (although its great to see it being used and finding issues so that we can fix them)
• Can you look at the samples in the solution - are they all identical or are they changing? Do trace plots work correctly?
• It looks like your model is quite far from representing your data - I would highly suggest modifying the parameters manually or with optimizers before using samplers. If your prior distributions are broad enough to cover the data, then this shouldn't be a problem, but you are using distributions sampled around your current face values, so its possible dynesty never sees enough improvement to actually move. In dynamic nested sampling, your priors should also be informative, so you shouldn't use gaussian priors unless you know that information in advance. You could switch out with broad uniform priors, but then it will likely become prohibitively slow - again, the best solution here is to use optimizers first which are better designed for finding the reasonable area of parameter space.

[mas0nr]

Thanks for your quick response Kyle!

To answer point 2, yes the samples change (also new values are suggested for adoption for each parameter in priors as expecteds) and attempting to plot in the styles 'trace', 'run' or 'failed' causes phoebe to raise:

TypeError                                 Traceback (most recent call last)
Cell In[20], line 1
----> 1 print(b['dynesty_run2'].plot(style='failed'))

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/parameters/parameters.py:5872, in ParameterSet.plot(self, twig, **kwargs)
   5870 else:
   5871     ps = self
-> 5872 plot_kwargss = ps._unpack_plotting_kwargs(animate=animate, **kwargs)
   5873 # print("*** plot_kwargss", plot_kwargss)
   5874 # this loop handles any of the automatically-generated
   5875 # multiple plotting calls, passing each on to autofig
   5876 for plot_kwargs in plot_kwargss:

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/parameters/parameters.py:4958, in ParameterSet._unpack_plotting_kwargs(self, animate, **kwargs)
   4955     raise ValueError("style must be a (single) string for dynesty")
   4957 if style != 'corner':
-> 4958     kwargs['results'] = _helpers.get_dynesty_object_from_solution(ps._bundle, ps.solution, adopt_parameters=kwargs.get('adopt_parameters'))
   4960 if style in ['corner', 'failed']:
   4961     kwargs['plot_package'] = 'distl'

File ~/miniconda3/envs/Summer_Research/lib/python3.12/site-packages/phoebe/helpers/__init__.py:405, in get_dynesty_object_from_solution(b, solution, adopt_parameters)
    401     raise ValueError("solution_ps must have kind 'dynesty'")
    403 adopt_inds, adopt_uniqueids = b._get_adopt_inds_uniqueids(solution_ps, adopt_parameters=adopt_parameters)
--> 405 return get_dynesty_object(adopt_inds=adopt_inds, **{p.qualifier: p.get_value() for p in solution_ps.to_list() if p.qualifier in ['nlive', 'niter', 'ncall', 'eff', 'samples', 'samples_id', 'samples_it', 'samples_u', 'logwt', 'logl', 'logvol', 'logz', 'logzerr', 'information', 'bound_iter', 'samples_bound', 'scale']})
...
    440                                 logwt=logwt, logl=logl, logvol=logvol, logz=logz, logzerr=logzerr,
    441                                 information=information, bound_iter=bound_iter,
    442                                 samples_bound=samples_bound, scale=scale)

TypeError: Results.__init__() got an unexpected keyword argument 'nlive'

as for point 3, I'm fairly confident that my initial values were obtained from an optimizer already but I'm going to rerun nelder-mead just in case. I'm no longer expecting to for this model to serve any use beyond my own learning having already spent a week looking for a local solution that produces reasonable masses which is why I had gone on to dynesty with such a poor fit. If dynesty won't function correctly as a result, that would be understandable and I will move onto another system that I have been tasked with modeling. Do you think that switching to broader uniform priors would allow me to get something useful out of dynesty?

[kecnry]

Do you think that switching to broader uniform priors would allow me to get something useful out of dynesty?

Its hard to say for sure, it definitely sounds like there is some sort of bug, but I'm not sure when we'll be able to track it down. The most tested/dependable workflow is to use optimizers to get to the point where the residuals are fairly symmetric and then using mcmc to get posteriors.

[mas0nr]

Thank you for your help Kyle! No rush on my issue, I've moved onto the next system (which is certainly eclipsing and should be possible to fit with physically sensible parameters) and I'll follow up if I happen to run into the same issues there.