Generating Bounded Model Inputs
===============================

After initial normalization, a feature can still be unbounded or strongly
skewed.  The map classes in this module define the transformations from the
normalized features to the final inputs used by the ML model
or mapped evaluator.  Each map records its raw input
indices and parameters, evaluates one transformed feature, and propagates
its derivative back to every raw input it uses.

``FeatureList`` stores these maps in the order expected from the model.  ``fill_vals_`` applies
the forward maps, ``fill_derivs_`` applies their adjoints, and YAML
serialization preserves map types and parameters.  Examples include the
rational map :math:`y=\gamma x/(1+\gamma x)` and the ``SLTMap`` used for the
bounded CIDER26XC kinetic-energy indicator.  Model loading supplies the exact
list associated with a packaged functional.

.. autoclass:: ciderpress.dft.transform_data.FeatureList

.. automodule:: ciderpress.dft.transform_data
   :members: