Number of steps cut with fixed parameters

[vandalt]

In plot_posterior, when checking the fixed parameter, only axis 0 has :, so axes 1 and 2 both get filtered by the index_sel mask. This results in each fitted parameter having only 4 steps per walker in the corner plot (the shape before the index selection is (nwalkers, nsteps-burnin, nparams) and it goes to (nwalkers, nparams, nparams) after). I fixed this in a fork of the project and I can create a PR (the only modification is shown in the code block below). I checked with a few numpy versions and it does not seem related to a numpy version problem.

    # Check if parameter values were fixed

    index_sel = []
    index_del = []

    # Use only last axis for parameter dimensions
    for i in range(ndim):
        if np.amin(samples[:, :, i]) == np.amax(samples[:, :, i]):
            index_del.append(i)
        else:
            index_sel.append(i)

    samples = samples[:, :, index_sel]

[tomasstolker]

Thanks for pointing out that issue @vandalt!

Right, this is something that I had implemented recently. The problem is that the ndim of the stored samples from run_mcmc is 3, while the ndim from run_multinest is 2. In the latter case the shape if (nsamples, nparams).

It would be great if you could create a PR which solves this issues and takes the two ndim cases into account.

By the way, when using FitModel, the run_multinest function has more features compared to run_mcmc and convergences more easily in case the posterior is multimodal. It also returns the evidence which can be used for model comparison.

[vandalt]

Right, I had not tried run_multinest so far so I did not think about that part, thank you.

I will create a PR that handles both cases. Ellipsis would probably be the best way to achieve this (i.e. samples[..., i]).

Thanks for the tip about run_multinest vs run_mcmc.