Find the scikit-learn trees behind a fitted imodels model.
Several features (rule extraction, leaf membership) need the same thing: the sklearn tree structures a model is built from, whatever shape the model takes. Keeping that in one place stops those features from disagreeing about which models are tree-based.
Expand source code
"""Find the scikit-learn trees behind a fitted imodels model.
Several features (rule extraction, leaf membership) need the same thing: the
sklearn tree structures a model is built from, whatever shape the model takes.
Keeping that in one place stops those features from disagreeing about which
models are tree-based.
"""
import numpy as np
#: attributes under which a model may keep the fitted model it delegates to.
#: `est_`/`best_estimator_`/`_final_estimator` cover the search-and-pipeline
#: chain that AutoInterpretableModel builds: est_ is a GridSearchCV whose
#: best_estimator_ is a Pipeline ending in the model actually chosen.
WRAPPED_MODEL_ATTRS = ('figs', 'estimator_', 'model',
'est_', 'best_estimator_', '_final_estimator')
def is_sklearn_tree(estimator) -> bool:
"""Whether `estimator` holds a fitted sklearn tree.
Checks the tree structure itself, since C45TreeClassifier also has a `tree_`
but stores XML in it.
"""
return hasattr(getattr(estimator, 'tree_', None), 'feature')
def n_tree_outputs(model) -> int:
"""Number of values each leaf of `model` holds (classes, or 1 for regression)."""
n_outputs = getattr(model, 'n_outputs', None)
if n_outputs:
return int(n_outputs)
classes = getattr(model, 'classes_', None)
return len(classes) if classes is not None else 1
def sklearn_trees(model, convert_figs=True):
"""The fitted sklearn trees making up `model`, or None if it has none.
Parameters
----------
model
A fitted model.
convert_figs : bool
Whether to convert a FIGS model's own tree objects into sklearn trees.
Set False when the caller reads FIGS nodes directly (the converted trees
store class counts rather than predictions).
"""
if hasattr(model, 'trees_'): # FIGS: a sum of its own tree objects
if not convert_figs:
return None
from imodels.tree.viz_utils import extract_sklearn_tree_from_figs
n_outputs = n_tree_outputs(model)
return [extract_sklearn_tree_from_figs(model, i, n_outputs)
for i in range(len(model.trees_))]
if is_sklearn_tree(model):
return [model]
# models that delegate to another fitted model (shrinkage, CV wrappers, TAO)
for attr in WRAPPED_MODEL_ATTRS:
inner = getattr(model, attr, None)
if inner is not None and inner is not model:
trees = sklearn_trees(inner, convert_figs=convert_figs)
if trees is not None:
return trees
subestimators = getattr(model, 'estimators_', None)
if subestimators is not None and len(subestimators) > 0:
trees = []
for estimator in subestimators:
if isinstance(estimator, np.ndarray): # gradient boosting nests them
estimator = estimator[0]
if not is_sklearn_tree(estimator):
return None
trees.append(estimator)
return trees
return NoneGlobal variables
var WRAPPED_MODEL_ATTRS-
attributes under which a model may keep the fitted model it delegates to.
est_/best_estimator_/_final_estimatorcover the search-and-pipeline chain that AutoInterpretableModel builds: est_ is a GridSearchCV whose best_estimator_ is a Pipeline ending in the model actually chosen.
Functions
def is_sklearn_tree(estimator) ‑> bool-
Whether
estimatorholds a fitted sklearn tree.Checks the tree structure itself, since C45TreeClassifier also has a
tree_but stores XML in it.Expand source code
def is_sklearn_tree(estimator) -> bool: """Whether `estimator` holds a fitted sklearn tree. Checks the tree structure itself, since C45TreeClassifier also has a `tree_` but stores XML in it. """ return hasattr(getattr(estimator, 'tree_', None), 'feature') def n_tree_outputs(model) ‑> int-
Number of values each leaf of
modelholds (classes, or 1 for regression).Expand source code
def n_tree_outputs(model) -> int: """Number of values each leaf of `model` holds (classes, or 1 for regression).""" n_outputs = getattr(model, 'n_outputs', None) if n_outputs: return int(n_outputs) classes = getattr(model, 'classes_', None) return len(classes) if classes is not None else 1 def sklearn_trees(model, convert_figs=True)-
The fitted sklearn trees making up
model, or None if it has none.Parameters
model- A fitted model.
convert_figs:bool- Whether to convert a FIGS model's own tree objects into sklearn trees. Set False when the caller reads FIGS nodes directly (the converted trees store class counts rather than predictions).
Expand source code
def sklearn_trees(model, convert_figs=True): """The fitted sklearn trees making up `model`, or None if it has none. Parameters ---------- model A fitted model. convert_figs : bool Whether to convert a FIGS model's own tree objects into sklearn trees. Set False when the caller reads FIGS nodes directly (the converted trees store class counts rather than predictions). """ if hasattr(model, 'trees_'): # FIGS: a sum of its own tree objects if not convert_figs: return None from imodels.tree.viz_utils import extract_sklearn_tree_from_figs n_outputs = n_tree_outputs(model) return [extract_sklearn_tree_from_figs(model, i, n_outputs) for i in range(len(model.trees_))] if is_sklearn_tree(model): return [model] # models that delegate to another fitted model (shrinkage, CV wrappers, TAO) for attr in WRAPPED_MODEL_ATTRS: inner = getattr(model, attr, None) if inner is not None and inner is not model: trees = sklearn_trees(inner, convert_figs=convert_figs) if trees is not None: return trees subestimators = getattr(model, 'estimators_', None) if subestimators is not None and len(subestimators) > 0: trees = [] for estimator in subestimators: if isinstance(estimator, np.ndarray): # gradient boosting nests them estimator = estimator[0] if not is_sklearn_tree(estimator): return None trees.append(estimator) return trees return None