encode vi and update to work with multiple RVs (#235)

* encode vi and update to work with multiple RVs

* add missing tests

Author: aloctavodia

env-dev.yml

@@ -3,7 +3,7 @@ channels:
   - conda-forge
   - defaults
 dependencies:
-  - pymc==5.24.0
+  - pymc>=5.24.0
   - numba
   - matplotlib
   - numpy

env.yml

@@ -3,7 +3,7 @@ channels:
   - conda-forge
   - defaults
 dependencies:
-  - pymc==5.24.0
+  - pymc>=5.24.0
   - numba
   - matplotlib
   - numpy

pymc_bart/pgbart.py

@@ -37,6 +37,7 @@
     get_idx_left_child,
     get_idx_right_child,
 )
+from pymc_bart.utils import _encode_vi
 
 
 class ParticleTree:
@@ -118,7 +119,7 @@ class PGBART(ArrayStepShared):
     default_blocked = False
     generates_stats = True
     stats_dtypes_shapes: dict[str, tuple[type, list]] = {
-        "variable_inclusion": (object, []),
+        "variable_inclusion": (int, []),
         "tune": (bool, []),
     }
 
@@ -335,6 +336,8 @@ def astep(self, _):
         if not self.tune:
             self.bart.all_trees.append(self.all_trees)
 
+        variable_inclusion = _encode_vi(variable_inclusion)
+
         stats = {"variable_inclusion": variable_inclusion, "tune": self.tune}
         return self.sum_trees, [stats]
 

pymc_bart/utils.py

@@ -7,6 +7,7 @@
 import matplotlib.pyplot as plt
 import numpy as np
 import numpy.typing as npt
+import pymc as pm
 import pytensor.tensor as pt
 from arviz_base import rcParams
 from arviz_stats.base import array_stats
@@ -674,48 +675,66 @@ def _smooth_mean(
     return x_data, y_data
 
 
-def get_variable_inclusion(idata, X, labels=None, to_kulprit=False):
+def get_variable_inclusion(idata, X, model=None, bart_var_name=None, labels=None, to_kulprit=False):
     """
     Get the normalized variable inclusion from BART model.
 
     Parameters
     ----------
     idata : InferenceData
-        InferenceData containing a collection of BART_trees in sample_stats group
+        InferenceData with a variable "variable_inclusion" in ``sample_stats`` group
     X : npt.NDArray
         The covariate matrix.
+    model : Optional[pm.Model]
+        The PyMC model that contains the BART variable. Only needed if the model contains multiple
+        BART variables.
+    bart_var_name : Optional[str]
+        The name of the BART variable in the model. Only needed if the model contains multiple
+        BART variables.
     labels : Optional[list[str]]
         List of the names of the covariates. If X is a DataFrame the names of the covariables will
         be taken from it and this argument will be ignored.
     to_kulprit : bool
         If True, the function will return a list of list with the variables names.
         This list can be passed as a path to Kulprit's project method. Defaults to False.
+
     Returns
     -------
     VI_norm : npt.NDArray
         Normalized variable inclusion.
     labels : list[str]
         List of the names of the covariates.
     """
-    VIs = idata["sample_stats"]["variable_inclusion"].mean(("chain", "draw")).values
+    n_vars = X.shape[1]
+    vi_xarray = idata["sample_stats"]["variable_inclusion"]
+    if "variable_inclusion_dim_0" in vi_xarray.coords:
+        if model is None or bart_var_name is None:
+            raise ValueError(
+                "The InfereceData was generated from a model with multiple BART variables, \n"
+                "please provide the model and also the name of the BART variable \n"
+                "for which you want to compute the variable inclusion."
+            )
+        index = [var.name for var in model.free_RVs].index(bart_var_name)
+        vi_vals = vi_xarray.sel({"variable_inclusion_dim_0": index}).values.ravel()
+    else:
+        vi_vals = idata["sample_stats"]["variable_inclusion"].values.ravel()
+    VIs = np.array([_decode_vi(val, n_vars) for val in vi_vals]).sum(axis=0)
     VI_norm = VIs / VIs.sum()
     idxs = np.argsort(VI_norm)
 
     indices = idxs[::-1]
     n_vars = len(indices)
 
     if hasattr(X, "columns") and hasattr(X, "to_numpy"):
-        labels = X.columns
+        labels = list(X.columns)
 
     if labels is None:
-        labels = np.arange(n_vars).astype(str)
-
-    label_list = labels.to_list()
+        labels = [str(i) for i in range(n_vars)]
 
     if to_kulprit:
-        return [label_list[:idx] for idx in range(n_vars)]
+        return [labels[:idx] for idx in range(n_vars)]
     else:
-        return VI_norm[indices], label_list
+        return VI_norm[indices], labels
 
 
 def plot_variable_inclusion(idata, X, labels=None, figsize=None, plot_kwargs=None, ax=None):
@@ -781,22 +800,26 @@ def compute_variable_importance(  # noqa: PLR0915 PLR0912
     idata: Any,
     bartrv: Variable,
     X: npt.NDArray,
+    model: "pm.Model | None" = None,
     method: str = "VI",
     fixed: int = 0,
     samples: int = 50,
-    random_seed: Optional[int] = None,
+    random_seed: int | None = None,
 ) -> dict[str, object]:
     """
     Estimates variable importance from the BART-posterior.
 
     Parameters
     ----------
     idata : InferenceData
-        InferenceData containing a collection of BART_trees in sample_stats group
+        InferenceData containing a "variable_inclusion" variable in the sample_stats group.
     bartrv : BART Random Variable
         BART variable once the model that include it has been fitted.
     X : npt.NDArray
         The covariate matrix.
+    model : Optional[pm.Model]
+        The PyMC model that contains the BART variable. Only needed if the model contains multiple
+        BART variables.
     method : str
         Method used to rank variables. Available options are "VI" (default), "backward"
         and "backward_VI".
@@ -825,6 +848,7 @@ def compute_variable_importance(  # noqa: PLR0915 PLR0912
     rng = np.random.default_rng(random_seed)
 
     all_trees = bartrv.owner.op.all_trees
+    bart_var_name = bartrv.name
 
     if bartrv.ndim == 1:  # type: ignore
         shape = 1
@@ -858,9 +882,20 @@ def compute_variable_importance(  # noqa: PLR0915 PLR0912
     )
 
     if method in ["VI", "backward_VI"]:
-        idxs = np.argsort(
-            idata["sample_stats"]["variable_inclusion"].mean(("chain", "draw")).values
-        )
+        vi_xarray = idata["sample_stats"]["variable_inclusion"]
+        if "variable_inclusion_dim_0" in vi_xarray.coords:
+            if model is None:
+                raise ValueError(
+                    "The InfereceData was generated from a model with multiple BART variables, \n"
+                    "please provide the model and also the name of the BART variable \n"
+                    "for which you want to compute the variable inclusion."
+                )
+
+            index = [var.name for var in model.free_RVs].index(bart_var_name)
+            vi_vals = vi_xarray.sel({"variable_inclusion_dim_0": index}).values.ravel()
+        else:
+            vi_vals = idata["sample_stats"]["variable_inclusion"].values.ravel()
+        idxs = np.argsort(np.array([_decode_vi(val, n_vars) for val in vi_vals]).sum(axis=0))
         subsets: list[list[int]] = [list(idxs[:-i]) for i in range(1, len(idxs))]
         subsets.append(None)  # type: ignore
 
@@ -1226,3 +1261,39 @@ def _plot_hdi(x, y, smooth, color, alpha, smooth_kwargs, ax):
 
     ax.fill_between(x_data, y_data[:, 0], y_data[:, 1], color=color, alpha=alpha)
     return ax
+
+
+def _decode_vi(n: int, length: int) -> list[int]:
+    """
+    Decode the variable inclusion from the BART model.
+    """
+    bits = bin(n)[2:]
+    vi_list: list[int] = []
+    i = 0
+    while len(vi_list) < length:
+        # Count prefix ones
+        prefix_len = 0
+        while bits[i] == "1":
+            prefix_len += 1
+            i += 1
+        i += 1  # skip the '0'
+        b = bits[i : i + prefix_len]
+        vi_list.append(int(b, 2))
+        i += prefix_len
+    return vi_list
+
+
+def _encode_vi(vec: npt.NDArray) -> int:
+    """
+    Encode variable inclusion vector into a single integer.
+
+    The encoding is done by converting each element of the vector into a binary string,
+    where each element contributes a prefix of '1's followed by a '0' and its binary representation.
+    The final result is the integer representation of the concatenated binary string.
+    """
+    bits = ""
+    for x in vec:
+        b = bin(x)[2:]
+        prefix = "1" * len(b) + "0"
+        bits += prefix + b
+    return int(bits, 2)

requirements.txt

@@ -1,4 +1,4 @@
-pymc==5.24.0
+pymc>=5.24.0
 arviz-stats[xarray]>=0.6.0
 numba
 matplotlib

tests/test_bart.py

@@ -1,11 +1,12 @@
 import numpy as np
 import pymc as pm
 import pytest
-from numpy.testing import assert_almost_equal, assert_array_equal
+from numpy.testing import assert_almost_equal
 from pymc.initial_point import make_initial_point_fn
 from pymc.logprob.basic import transformed_conditional_logp
 
 import pymc_bart as pmb
+from pymc_bart.utils import _decode_vi
 
 
 def assert_moment_is_expected(model, expected, check_finite_logp=True):
@@ -52,14 +53,12 @@ def test_bart_vi(response):
     with pm.Model() as model:
         mu = pmb.BART("mu", X, Y, m=10, response=response)
         sigma = pm.HalfNormal("sigma", 1)
-        y = pm.Normal("y", mu, sigma, observed=Y)
+        pm.Normal("y", mu, sigma, observed=Y)
         idata = pm.sample(tune=200, draws=200, random_seed=3415)
-        var_imp = (
-            idata.sample_stats["variable_inclusion"]
-            .stack(samples=("chain", "draw"))
-            .mean("samples")
-        )
-        var_imp /= var_imp.sum()
+        vi_vals = idata["sample_stats"]["variable_inclusion"].values.ravel()
+        var_imp = np.array([_decode_vi(val, 3) for val in vi_vals]).sum(axis=0)
+
+        var_imp = var_imp / var_imp.sum()
         assert var_imp[0] > var_imp[1:].sum()
         assert_almost_equal(var_imp.sum(), 1)
 
@@ -123,92 +122,6 @@ def test_shape(response):
     assert idata.posterior.coords["w_dim_1"].data.size == 250
 
 
-class TestUtils:
-    X_norm = np.random.normal(0, 1, size=(50, 2))
-    X_binom = np.random.binomial(1, 0.5, size=(50, 1))
-    X = np.hstack([X_norm, X_binom])
-    Y = np.random.normal(0, 1, size=50)
-
-    with pm.Model() as model:
-        mu = pmb.BART("mu", X, Y, m=10)
-        sigma = pm.HalfNormal("sigma", 1)
-        y = pm.Normal("y", mu, sigma, observed=Y)
-        idata = pm.sample(tune=200, draws=200, random_seed=3415)
-
-    def test_sample_posterior(self):
-        all_trees = self.mu.owner.op.all_trees
-        rng = np.random.default_rng(3)
-        pred_all = pmb.utils._sample_posterior(all_trees, X=self.X, rng=rng, size=2)
-        rng = np.random.default_rng(3)
-        pred_first = pmb.utils._sample_posterior(all_trees, X=self.X[:10], rng=rng)
-
-        assert_almost_equal(pred_first[0], pred_all[0, :10], decimal=4)
-        assert pred_all.shape == (2, 50, 1)
-        assert pred_first.shape == (1, 10, 1)
-
-    @pytest.mark.parametrize(
-        "kwargs",
-        [
-            {},
-            {
-                "samples": 2,
-                "var_discrete": [3],
-            },
-            {"instances": 2},
-            {"var_idx": [0], "smooth": False, "color": "k"},
-            {"grid": (1, 2), "sharey": "none", "alpha": 1},
-            {"var_discrete": [0]},
-        ],
-    )
-    def test_ice(self, kwargs):
-        pmb.plot_ice(self.mu, X=self.X, Y=self.Y, **kwargs)
-
-    @pytest.mark.parametrize(
-        "kwargs",
-        [
-            {},
-            {
-                "samples": 2,
-                "xs_interval": "quantiles",
-                "xs_values": [0.25, 0.5, 0.75],
-                "var_discrete": [3],
-            },
-            {"var_idx": [0], "smooth": False, "color": "k"},
-            {"grid": (1, 2), "sharey": "none", "alpha": 1},
-            {"var_discrete": [0]},
-        ],
-    )
-    def test_pdp(self, kwargs):
-        pmb.plot_pdp(self.mu, X=self.X, Y=self.Y, **kwargs)
-
-    @pytest.mark.parametrize(
-        "kwargs",
-        [
-            {"samples": 50},
-            {"labels": ["A", "B", "C"], "samples": 2, "figsize": (6, 6)},
-        ],
-    )
-    def test_vi(self, kwargs):
-        samples = kwargs.pop("samples")
-        vi_results = pmb.compute_variable_importance(
-            self.idata, bartrv=self.mu, X=self.X, samples=samples
-        )
-        pmb.plot_variable_importance(vi_results, **kwargs)
-        pmb.plot_scatter_submodels(vi_results, **kwargs)
-
-    def test_pdp_pandas_labels(self):
-        pd = pytest.importorskip("pandas")
-
-        X_names = ["norm1", "norm2", "binom"]
-        X_pd = pd.DataFrame(self.X, columns=X_names)
-        Y_pd = pd.Series(self.Y, name="response")
-        axes = pmb.plot_pdp(self.mu, X=X_pd, Y=Y_pd)
-
-        figure = axes[0].figure
-        assert figure.texts[0].get_text() == "Partial response"
-        assert_array_equal([ax.get_xlabel() for ax in axes], X_names)
-
-
 @pytest.mark.parametrize(
     "size, expected",
     [
@@ -275,7 +188,7 @@ def test_multiple_bart_variables():
 
         # Combined model
         sigma = pm.HalfNormal("sigma", 1)
-        y = pm.Normal("y", mu1 + mu2, sigma, observed=Y)
+        pm.Normal("y", mu1 + mu2, sigma, observed=Y)
 
         # Sample with automatic assignment of BART samplers
         idata = pm.sample(tune=50, draws=50, chains=1, random_seed=3415)
@@ -291,6 +204,16 @@ def test_multiple_bart_variables():
         assert idata.posterior["mu1"].shape == (1, 50, 50)
         assert idata.posterior["mu2"].shape == (1, 50, 50)
 
+        vi_results = pmb.compute_variable_importance(idata, mu1, X1, model=model)
+        assert vi_results["labels"].shape == (2,)
+        assert vi_results["preds"].shape == (2, 50, 50)
+        assert vi_results["preds_all"].shape == (50, 50)
+
+        vi_tuple = pmb.get_variable_inclusion(idata, X1, model=model, bart_var_name="mu1")
+        assert vi_tuple[0].shape == (2,)
+        assert len(vi_tuple[1]) == 2
+        assert isinstance(vi_tuple[1][0], str)
+
 
 def test_multiple_bart_variables_manual_step():
     """Test that multiple BART variables work with manually assigned PGBART samplers."""