Hssm

This version of HSSM incorporated the following changes:

1. We updated HSSM to be compatible with the major API changes in `bambi` v0.14.0.
2. We fixed various graphing issues in `pymc` 5.16.0+, thanks to the API changes in `bambi`.
3. We added variational inference via native `pymc`.
4. We can now use `float64` inference.
5. We fixed some minor bugs in providing initial values.
6. We added a model.dic() convenience function.
7. We added a model.restore_traces() convenience function.
8. Other minor bug fixes.

This is a maintenance release of HSSM, mainly to add a version constraint on `bambi` in light of the many breaking changes that version `0.1.4` introduces. This version also improved compatibility with `PyMC>=5.15` and incorporated minor bug fixes:

1. We incorporated a temporary fix to graphing which broke after `PyMC>=5.15`.
2. We deprecated `ndim` and `ndim_supp` definition in `SSMRandomVariable` in `PyMC>-5.16`.
3. We fixed a bug that prevents new traces from being returned if `model.sample()` is called again.

HSSM is now on Conda! We now recommend installing HSSM through `conda install -c conda-forge hssm`. For advanced users, we also support installing the GPU version of JAX through `pip install hssm[cuda12]`.

This version incorporates various bug fixes:

1. We fixed a major bug that causes divergences for models using `approx_differentiable` and `blackbox` likelihoods. We are still looking into the issues of divergence with `analytical` likelihoods.
2. We made the model creation process more robust, fixing errors when categorical variables are used with group identifiers.
3. We updated the codebase according to the deprecations in higher versions of JAX.
4. We implemented a temporary fix to an issue that might cause the kernel to die due to OOM.

We added a few new features in 0.2.1:

1. We have finished updating the HSSM code base to support go-nogo data and deadline. We will provide documentation once the networks are added to our huggingface repo.
2. We updated `hssm.distribution_utils` to streamline the creation of `pm.Distribution`s.
3. We now support response variables other than `rt` and `response`. They can be specified through `model_config` via the new `response` field.
4. We have fixed some of the issues with convergence when using `log-logit` link functions and/or safe priors.

Other minor updates

- Fixed an incompatible shape error during posterior predictive sampling when `p_outlier` is estimated as a parameter.
- Updated documentation for using `make_distribution` with PyMC.

Bug fixes:

- Fixed default list of parameters for `ddm_full` model and the bounds for `ddm_sdv` model.

This is a major version update! Many changes have taken place in this version:

Breaking changes

When `hierarchical` argument of `hssm.HSSM` is set to `True`, HSSM will look into the
`data` provided for the `participant_id` field. If it does not exist, an error will
be thrown.

New features

- Added `link_settings` and `prior_settings` arguments to `hssm.HSSM`, which allows HSSM
to use intelligent default priors and link functions for complex hierarchical models.

- Added an `hssm.plotting` submodule with `plot_posterior_predictive()` and
`plot_quantile_probability` for creating posterior predictive plots and quantile
probability plots.

- Added an `extra_fields` argument to `hssm.HSSM` to pass additional data to the
likelihood function computation.

- Limited `PyMC`, `pytensor`, `numpy`, and `jax` dependency versions for compatibility.

0.1.x

We fixed the errors in v0.1.4. Sorry for the convenience! If you have accidentally
downloaded v0.1.4, please make sure that you update hssm to the current version.

- We made Cython dependencies of this package available via pypi. We have also built
wheels for (almost) all platforms so there is no need to build these Cython
dependencies.