Reset ensemble coefficients before each fit

pysindy/optimizers/base.py

@@ -385,6 +385,7 @@ def __init__(
         self.unbias = False
 
     def _reduce(self, x: AxesArray, y: np.ndarray) -> None:
+        self.coef_list = []
         x = AxesArray(np.asarray(x), {"ax_sample": 0, "ax_coord": 1})
         n_samples = x.shape[x.ax_sample]
         if self.bagging and self.n_subset is None:

test/test_optimizers/test_ensemble_refit.py

@@ -0,0 +1,111 @@
+import numpy as np
+import pytest
+
+from pysindy import SINDy
+from pysindy.feature_library import IdentityLibrary
+from pysindy.optimizers import EnsembleOptimizer
+from pysindy.optimizers import STLSQ
+
+
+def mean_members(members):
+    return np.mean(members, axis=0)
+
+
+def full_sample_ensemble(n_samples, **kwargs):
+    return EnsembleOptimizer(
+        STLSQ(threshold=0, alpha=0),
+        bagging=True,
+        n_models=3,
+        n_subset=n_samples,
+        replace=False,
+        **kwargs,
+    )
+
+
+@pytest.mark.parametrize("aggregator", [None, mean_members], ids=["median", "mean"])
+def test_refit_uses_only_latest_training_data(aggregator):
+    # Every bag contains every row. Permuting the rows cannot change the
+    # unique least-squares solution for these full-rank designs.
+    x = np.tile(np.eye(2), (3, 1))
+    initial = np.array([[1.0, -2.0]])
+    current = np.array([[3.0, 4.0]])
+    optimizer = full_sample_ensemble(len(x), ensemble_aggregator=aggregator)
+
+    optimizer.fit(x, x @ initial.T)
+    np.testing.assert_allclose(optimizer.coef_, initial, rtol=1e-12, atol=1e-12)
+    optimizer.fit(x, x @ current.T)
+
+    np.testing.assert_allclose(optimizer.coef_, current, rtol=1e-12, atol=1e-12)
+    assert len(optimizer.coef_list) == 3
+    for member in optimizer.coef_list:
+        np.testing.assert_allclose(member, current, rtol=1e-12, atol=1e-12)
+
+
+def test_refit_allows_different_coefficient_shapes():
+    initial_x = np.tile(np.eye(2), (3, 1))
+    current_x = np.tile(np.eye(3), (2, 1))
+    initial = np.array([[1.0, -2.0]])
+    current = np.array([[3.0, 0.0, -1.0], [0.0, 4.0, 2.0]])
+    optimizer = full_sample_ensemble(len(initial_x))
+
+    optimizer.fit(initial_x, initial_x @ initial.T)
+    optimizer.fit(current_x, current_x @ current.T)
+
+    assert optimizer.coef_.shape == current.shape
+    np.testing.assert_allclose(optimizer.coef_, current, rtol=1e-12, atol=1e-12)
+    assert len(optimizer.coef_list) == 3
+    for member in optimizer.coef_list:
+        assert member.shape == current.shape
+        np.testing.assert_allclose(member, current, rtol=1e-12, atol=1e-12)
+
+
+@pytest.mark.parametrize(
+    "modes",
+    [
+        {"bagging": True},
+        {"library_ensemble": True},
+        {"bagging": True, "library_ensemble": True},
+    ],
+    ids=["bagging", "library", "both"],
+)
+def test_refit_updates_member_count_and_preserves_previous_snapshot(modes):
+    # The zero response has zero coefficients for every possible library
+    # subset, so this tests member ownership without depending on random draws.
+    x = np.tile(np.eye(3), (2, 1))
+    y = np.zeros((len(x), 1))
+    optimizer = EnsembleOptimizer(
+        STLSQ(threshold=0, alpha=0),
+        n_models=3,
+        n_subset=len(x),
+        replace=False,
+        **modes,
+    )
+    optimizer.fit(x, y)
+    previous_members = optimizer.coef_list
+    previous_values = [member.copy() for member in previous_members]
+    assert len(previous_members) == 3
+
+    optimizer.set_params(n_models=2)
+    optimizer.fit(x, y)
+
+    assert len(optimizer.coef_list) == 2
+    assert optimizer.coef_list is not previous_members
+    np.testing.assert_array_equal(optimizer.coef_, np.zeros((1, 3)))
+    assert len(previous_members) == 3
+    for retained, expected in zip(previous_members, previous_values):
+        np.testing.assert_array_equal(retained, expected)
+
+
+def test_repeated_sindy_fit_uses_latest_coefficients():
+    x = np.tile(np.eye(2), (3, 1))
+    initial = np.array([[1.0, 0.0], [0.0, -2.0]])
+    current = np.array([[3.0, -1.0], [2.0, 4.0]])
+    optimizer = full_sample_ensemble(len(x))
+    model = SINDy(feature_library=IdentityLibrary(), optimizer=optimizer)
+
+    model.fit(x, t=1.0, x_dot=x @ initial.T)
+    np.testing.assert_allclose(model.coefficients(), initial, rtol=1e-12, atol=1e-12)
+    model.fit(x, t=1.0, x_dot=x @ current.T)
+
+    np.testing.assert_allclose(model.coefficients(), current, rtol=1e-12, atol=1e-12)
+    assert len(optimizer.coef_list) == 3