Checkpointing is scrambling parameter labels.

[sharanbngr]

I've run into a strange error where checkpointing is scrambling the parameter labels. Trace plots are produced correctly and the final result file looks sensible as long as bilby is not loading a checkpoint during its run. This is NOT happening all the time but only with this particular model / parameter labels. My hunch is that my parameter names are probably breaking some hidden rule

In one of the runs where I did load from a checkpoint, most of the posterior samples of this parameter are outside its prior range. And the even more bizarre thing is that the every time I read in the checkpoint file, I get a new random permutation of the label-sample association.

(Pdb) priors["sigma_p_0"]
Uniform(minimum=0.01, maximum=1.0, name='sigma_p_0', latex_label='$\\sigma^0_{p}$', unit=None, boundary=None)
(Pdb) result.posterior['sigma_p_0'].max()
4.862758644802147

For reference, here are the list of parameter labels I'm using. I'm using dynesty.

(Pdb) priors.keys()
dict_keys(['mb_0', 'beta_0', 'alpha', 'sigpp_2', 'lam_0', 'lamb_0', 'sigpp_1', 'lam_1', 'delta_m_2', 'mpp_1', 'mmax', 'sigma_eff_0', 'beta_1', 'sigma_p_0', 'mu_p_0', 'mpp_2', 'mlow_2', 'delta_m_1', 'mu_eff_0', 'mlow_1'])

[fgittins]

Thanks for raising this. Are you able to put together a minimal reproducible example for this issue?

[sharanbngr]

Will try to get back to you in a couple of days with a minimal example.