bugfix for fit_laplace absent dims

pymc_extras/inference/laplace_approx/laplace.py

@@ -14,6 +14,7 @@
 
 
 import logging
+import re
 
 from collections.abc import Callable
 from functools import partial
@@ -51,6 +52,58 @@
 _log = logging.getLogger(__name__)
 
 
+def _reset_laplace_dim_idx(idata: az.InferenceData) -> az.InferenceData:
+    """
+    Because `fit_laplace` adds the (temp_chain, temp_draw) dimensions,
+    any variables without explicitly assigned dimensions receive
+    automatically generated indices that are shifted by two during
+    InferenceData creation.
+
+    This helper function corrects that shift by subtracting 2 from the
+    automatically detected dimension indices of the form
+    `<varname>_dim_<idx>`, restoring them to the indices they would have
+    had if the (temp_chain, temp_draw) dimensions were not added.
+
+    Only affects auto-assigned dimensions in `idata.posterior`.
+    """
+
+    pattern = re.compile(r"^(?P<base>.+)_dim_(?P<idx>\d+)$")
+
+    dim_renames = {}
+    var_renames = {}
+
+    for dim in idata.posterior.dims:
+        match = pattern.match(dim)
+        if match is None:
+            continue
+
+        base = match.group("base")
+        idx = int(match.group("idx"))
+
+        # Guard against invalid or unintended renames
+        if idx < 2:
+            raise ValueError(
+                f"Cannot reset Laplace dimension index for '{dim}': "
+                f"index {idx} would become negative."
+            )
+
+        new_dim = f"{base}_dim_{idx - 2}"
+
+        dim_renames[dim] = new_dim
+
+        # Only rename variables if they actually exist
+        if dim in idata.posterior.variables:
+            var_renames[dim] = new_dim
+
+    if dim_renames:
+        idata.posterior = idata.posterior.rename_dims(dim_renames)
+
+    if var_renames:
+        idata.posterior = idata.posterior.rename_vars(var_renames)
+
+    return idata
+
+
 def get_conditional_gaussian_approximation(
     x: TensorVariable,
     Q: TensorVariable | ArrayLike,
@@ -224,12 +277,8 @@ def model_to_laplace_approx(
             elif name in model.named_vars_to_dims:
                 dims = (*batch_dims, *model.named_vars_to_dims[name])
             else:
-                dims = (*batch_dims, *[f"{name}_dim_{i}" for i in range(batched_rv.ndim - 2)])
-                initval = initial_point.get(name, None)
-                dim_shapes = initval.shape if initval is not None else batched_rv.type.shape[2:]
-                laplace_model.add_coords(
-                    {name: np.arange(shape) for name, shape in zip(dims[2:], dim_shapes)}
-                )
+                n_dim = batched_rv.ndim - 2  # (temp_chain, temp_draw) are always first 2 dims
+                dims = (*batch_dims,) + (None,) * n_dim
 
             pm.Deterministic(name, batched_rv, dims=dims)
 
@@ -468,4 +517,6 @@ def fit_laplace(
             ["laplace_approximation", "unpacked_variable_names"]
         )
 
+        idata = _reset_laplace_dim_idx(idata)
+
     return idata

tests/inference/laplace_approx/test_laplace.py

@@ -193,6 +193,44 @@ def test_fit_laplace_ragged_coords(rng):
     assert (idata["posterior"].beta.sel(feature=1).to_numpy() > 0).all()
 
 
+def test_fit_laplace_no_data_or_deterministic_dims(rng):
+    coords = {"city": ["A", "B", "C"], "feature": [0, 1], "obs_idx": np.arange(100)}
+    with pm.Model(coords=coords) as ragged_dim_model:
+        X = pm.Data("X", np.ones((100, 2)))
+        beta = pm.Normal(
+            "beta", mu=[[-100.0, 100.0], [-100.0, 100.0], [-100.0, 100.0]], dims=["city", "feature"]
+        )
+        mu = pm.Deterministic("mu", (X[:, None, :] * beta[None]).sum(axis=-1))
+        sigma = pm.Normal("sigma", mu=1.5, sigma=0.5, dims=["city"])
+
+        obs = pm.Normal(
+            "obs",
+            mu=mu,
+            sigma=sigma,
+            observed=rng.normal(loc=3, scale=1.5, size=(100, 3)),
+            dims=["obs_idx", "city"],
+        )
+
+        idata = fit_laplace(
+            optimize_method="Newton-CG",
+            progressbar=False,
+            use_grad=True,
+            use_hessp=True,
+        )
+
+    # These should have been dropped when the laplace idata was created
+    assert "laplace_approximation" not in list(idata.posterior.data_vars.keys())
+    assert "unpacked_var_names" not in list(idata.posterior.coords.keys())
+
+    assert idata["posterior"].beta.shape[-2:] == (3, 2)
+    assert idata["posterior"].sigma.shape[-1:] == (3,)
+
+    # Check that everything got unraveled correctly -- feature 0 should be strictly negative, feature 1
+    # strictly positive
+    assert (idata["posterior"].beta.sel(feature=0).to_numpy() < 0).all()
+    assert (idata["posterior"].beta.sel(feature=1).to_numpy() > 0).all()
+
+
 def test_model_with_nonstandard_dimensionality(rng):
     y_obs = np.concatenate(
         [rng.normal(-1, 2, size=150), rng.normal(3, 1, size=350), rng.normal(5, 4, size=50)]