KarlNaumann / MacroStat / 24153505031

91%

push

github

fix(NK3E): preserve autograd graph in initialize and align with conventions (#68)

* fix(core): filter hyperparameter overrides by hyper keys in Parameters

Parameters.__init__ filtered user-supplied hyperparameters against
self.values (parameter names) instead of self.hyper (hyperparameter
names), silently dropping all overrides. Fix: self.values -> self.hyper.

* fix(NK3E): preserve autograd graph in initialize, use base forward

Replace torch.tensor([...]) with .reshape(1) in initialize() to keep
the computation graph from nn.Parameters to state variables. Remove
custom forward() override — NK3E now uses the inherited Behavior.forward()
like all other models. Remove unused tqdm import.

* test(NK3E): add autograd verification and refactor helpers

Add test_initialize_preserves_autograd_graph verifying dr/da1 matches
the analytical derivative. Extract _make_model helper. Update timing
comments for base class forward convention.

* docs(NK3E): rewrite notebook to match GL06 documentation template

Replace minimal 5-cell notebook with full structured documentation:
model intro with source references, module contents toctree, behavioral
equations in LaTeX, implementation notes, baseline convergence plot,
and three dedicated perturbation sections with comparative plots.
Add jupytext .py pair.

406 of 415 branches covered (97.83%)

Branch coverage included in aggregate %.

1 of 1 new or added line in 1 file covered. (100.0%)

2819 of 2954 relevant lines covered (95.43%)

0.95 hits per line