GitLab

The shifting and scaling was not being applied consistently to the
beta distribution's PDF, and values outside the supported range were being
extrapolated.  This had no impact for distributions on the standard
[0, 1] scale, as no shifting or scaling was needed, and in the case of
spin distributions, values outside the range were never passed in.

We now correctly apply the shifting and scaling, and perform a check
to ensure values are outside the range give a PDF of zero.  This will
impact performance, as extra operations are added, but will ensure the
output is always correct.  In the future we may want to add an option
to disable the bounds check for models using the [0, 1] scale, as it
isn't necessary there.

Broadcasting was broken for the beta and truncnorm models when the
parameters' dimensions exceed 1.  This occurs when `where=True`, as
`x[True].shape == (1,)+x.shape`, whereas the desired behavior is to
maintain the original shape.  To fix this, we simply broadcast the
`where` array to `parameters.shape`.

- Added a general framework for defining bins for arbitrary uses.
- Added a routine for computing rates restricted to said arbitrary bins.
- Added an executable for computing binned rates for the pl+g model

Added a truncated Gaussian prior, as well as a KDE prior.  KDE prior
loads samples from a specified file, and can optionally be
interpolated to reduce computational cost.  Neither is tested yet, so
no guarantees anything works.

When combining `--unique` and `--max-samples`, the potentially reduced
number of samples from limiting to unique samples causes the random
selection to satisfy `--max-samples` to potentially attempt to index
outside the array of remaining samples.

Now we step through the samples in order, terminating once enough
samples have posterior support to initialize all walkers.  This has
the side effect of removing the shuffling that was previously done.
If your input samples are not independent, be sure to shuffle them
ahead of time, e.g., with `pop_models_h5_cleaned_samples post_cleaned.hdf5 copy`
and `--max-samples` set less than the total.

Changed datatype from a list to a dict, but didn't change the for loop
until now.

Removing a copy+pasted plus sign

Rather than naively draw mean and variance values from an enclosing square, we limit ourselves to an enclosing triangle.  At the moment this actually excludes a small region of valid parameters, which should be filled in once the MCMC runs for a bit, but a future patch should fix this nonetheless.

The `copy` subcommand of `pop_models_h5_cleaned_samples` now has an
optional `--unique` flag which discards any duplicate posterior
samples.  This is useful when dealing with runs with stuck MCMC
walkers, particularly when using such a run to seed another run in
hopes of becoming un-stuck.

Previous dependency list did not allow h5py 3.0 or above as it was not
released and could break compatibility.  Now that it is out and does
not appear to break anything, released prior to the future 4.0 release
are allowed.

Specifically for the PL+G model, there is now a tool for computing the population-weighted average of a single event's likelihood.