v16.0.0

What's Changed

• Return dict of list of floats or list of floats in fitting functions by @danielward27 in #185
• No need for static by @danielward27 in #186
• Support old style keys in numpyro wrappers by @danielward27 in #187
• key_based_loss update by @danielward27 in #188
• Reduce coupling by @danielward27 in #189

Breaking changes:

• Calls to get_ravelled_pytree_constructor will now need to explicitly pass the *args and **kwargs for partitioning parameters (usually setting is_leaf=lambda leaf: isinstance(leaf, wrappers.NonTrainable).
• fit_to_data now returns list of floats, rather than than a list of scalar arrays.

Note, fit_to_variational_target will be deprecated in the next version. This version adds its replacement fit_to_key_based_loss. This was primarily because of some defaults which were "bad", e.g. steps: int = 100, and return_best=True (see #188 for details). It also generalizes the name, as it can be used to fit any pytree, and doesn't have to be used with a variational inference loss function.

Full Changelog: v15.1.0...v16.0.0