ngboost 0.5.9__tar.gz → 0.5.10.dev0__tar.gz

{ngboost-0.5.9 → ngboost-0.5.10.dev0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: ngboost
-Version: 0.5.9
+Version: 0.5.10.dev0
 Summary: Library for probabilistic predictions via gradient boosting.
 License: Apache-2.0
 Author: Stanford ML Group

{ngboost-0.5.9 → ngboost-0.5.10.dev0}/ngboost/__init__.py

@@ -7,8 +7,15 @@ except ImportError:
     from importlib_metadata import version
 
 from .api import NGBClassifier, NGBRegressor, NGBSurvival
+from .helpers import load_ngboost_model
 from .ngboost import NGBoost
 
-__all__ = ["NGBClassifier", "NGBRegressor", "NGBSurvival", "NGBoost"]
+__all__ = [
+    "NGBClassifier",
+    "NGBRegressor",
+    "NGBSurvival",
+    "NGBoost",
+    "load_ngboost_model",
+]
 
 __version__ = version(__name__)

ngboost-0.5.10.dev0/ngboost/helpers.py

@@ -0,0 +1,137 @@
+import sys
+import threading
+import types as _types
+
+import joblib
+import numpy as np
+import sklearn.tree._tree as _sklearn_tree  # pylint: disable=c-extension-no-member
+from sklearn.utils import check_array
+
+# ---------------------------------------------------------------------------
+# Backward compatibility helpers (issue #389)
+# ---------------------------------------------------------------------------
+
+_TREE_MODULE_SWAP_LOCK = threading.RLock()
+
+
+class _CompatTree(_sklearn_tree.Tree):  # pylint: disable=too-few-public-methods
+    """Transient subclass of sklearn's Tree used only during loading.
+
+    Overrides ``__setstate__`` to transparently inject the ``missing_go_to_left``
+    field that was added in scikit-learn 1.3, so that models pickled with older
+    sklearn can be loaded without raising ``ValueError``.
+    """
+
+    def __setstate__(self, state):
+        nodes = state.get("nodes")
+        if (
+            nodes is not None
+            and nodes.dtype.names is not None
+            and "missing_go_to_left" not in nodes.dtype.names
+        ):
+            new_dtype = np.dtype(nodes.dtype.descr + [("missing_go_to_left", "u1")])
+            new_nodes = np.zeros(nodes.shape, dtype=new_dtype)
+            for name in nodes.dtype.names:
+                new_nodes[name] = nodes[name]
+            new_nodes["missing_go_to_left"] = 1  # default: send missing values left
+            state = {**state, "nodes": new_nodes}
+        super().__setstate__(state)
+
+
+def _make_compat_tree_module():
+    """Return a drop-in replacement for sklearn.tree._tree with Tree → _CompatTree."""
+    mod = _types.ModuleType("sklearn.tree._tree")
+    for _attr in dir(_sklearn_tree):
+        setattr(mod, _attr, getattr(_sklearn_tree, _attr))
+    mod.Tree = _CompatTree
+    return mod
+
+
+def _to_proper_tree(compat_tree):
+    """Re-wrap a _CompatTree as a standard sklearn Tree (same state, proper type)."""
+    state = compat_tree.__getstate__()
+    n_classes = compat_tree.n_classes
+    if isinstance(n_classes, int):
+        n_classes = np.array([n_classes], dtype=np.intp)
+    proper = _sklearn_tree.Tree(  # pylint: disable=c-extension-no-member
+        compat_tree.n_features, n_classes.copy(), compat_tree.n_outputs
+    )
+    proper.__setstate__(state)
+    return proper
+
+
+def _fix_compat_trees(model):
+    """Replace every _CompatTree in model.base_models with a proper sklearn Tree."""
+    if not hasattr(model, "base_models"):
+        return
+    for iter_models in model.base_models:
+        for estimator in iter_models:
+            if hasattr(estimator, "tree_") and isinstance(estimator.tree_, _CompatTree):
+                estimator.tree_ = _to_proper_tree(estimator.tree_)
+
+
+def load_ngboost_model(filepath):
+    """Load an NGBoost model with backward compatibility for older scikit-learn versions.
+
+    Scikit-learn 1.3 added a ``missing_go_to_left`` field to the internal node
+    structure of decision trees.  Models trained and saved with scikit-learn < 1.3
+    do not contain this field, so loading them under scikit-learn >= 1.3 raises::
+
+        ValueError: node array from the pickle has an incompatible dtype
+
+    This function handles the incompatibility transparently by temporarily replacing
+    ``sklearn.tree._tree.Tree`` in ``sys.modules`` with a compatible subclass that
+    injects the missing field during ``__setstate__``.  After loading, all trees are
+    converted back to standard sklearn ``Tree`` objects so that subsequent use and
+    re-serialisation behave normally.
+
+    Args:
+        filepath: Path to the saved model file (joblib or pickle format).
+
+    Returns:
+        The loaded NGBoost model.
+
+    Example::
+
+        from ngboost import load_ngboost_model
+        model = load_ngboost_model("my_old_model.pkl")
+        preds = model.predict(X)
+    """
+    with _TREE_MODULE_SWAP_LOCK:
+        compat_module = _make_compat_tree_module()
+        original_module = sys.modules.get("sklearn.tree._tree")
+        sys.modules["sklearn.tree._tree"] = compat_module
+        try:
+            model = joblib.load(filepath)
+        finally:
+            if original_module is None:
+                sys.modules.pop("sklearn.tree._tree", None)
+            else:
+                sys.modules["sklearn.tree._tree"] = original_module
+
+    _fix_compat_trees(model)
+    return model
+
+
+# ---------------------------------------------------------------------------
+
+
+def Y_from_censored(T, E=None):
+    if T is None:
+        return None
+    if T.dtype == [
+        ("Event", "?"),
+        ("Time", "<f8"),
+    ]:  # already processed. Necessary for when d_score() calls score() as in LogNormalCRPScore
+        return T
+    T = check_array(T, ensure_2d=False)
+    T = T.reshape(T.shape[0])
+    if E is None:
+        E = np.ones_like(T)
+    else:
+        E = check_array(E, ensure_2d=False)
+        E = E.reshape(E.shape[0])
+    Y = np.empty(dtype=[("Event", np.bool_), ("Time", np.float64)], shape=T.shape[0])
+    Y["Event"] = E.astype(np.bool_)
+    Y["Time"] = T.astype(np.float64)
+    return Y

{ngboost-0.5.9 → ngboost-0.5.10.dev0}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "ngboost"
-version = "0.5.9"
+version = "0.5.10dev"
 description = "Library for probabilistic predictions via gradient boosting."
 authors = ["Stanford ML Group <avati@cs.stanford.edu>"]
 readme = "README.md"

ngboost-0.5.9/ngboost/helpers.py

@@ -1,23 +0,0 @@
-import numpy as np
-from sklearn.utils import check_array
-
-
-def Y_from_censored(T, E=None):
-    if T is None:
-        return None
-    if T.dtype == [
-        ("Event", "?"),
-        ("Time", "<f8"),
-    ]:  # already processed. Necessary for when d_score() calls score() as in LogNormalCRPScore
-        return T
-    T = check_array(T, ensure_2d=False)
-    T = T.reshape(T.shape[0])
-    if E is None:
-        E = np.ones_like(T)
-    else:
-        E = check_array(E, ensure_2d=False)
-        E = E.reshape(E.shape[0])
-    Y = np.empty(dtype=[("Event", np.bool_), ("Time", np.float64)], shape=T.shape[0])
-    Y["Event"] = E.astype(np.bool_)
-    Y["Time"] = T.astype(np.float64)
-    return Y