Provide helper method to initiliaze a MetaLearner based on another …

…`MetaLearner` (#71) * Provide helper method to initialize MetaLearner. * Fix logic revolving around pre-fitted models. * Add changelog entry. * Expand on docstring. * Compare attributes. * Update metalearners/metalearner.py Co-authored-by: Francesc Martí Escofet <[email protected]> * Update metalearners/metalearner.py Co-authored-by: Francesc Martí Escofet <[email protected]> --------- Co-authored-by: Francesc Martí Escofet <[email protected]>
Quantco · Jul 25, 2024 · 3cad00e · 3cad00e
1 parent 2f428fb
commit 3cad00e
Show file tree

Hide file tree

Showing 4 changed files with 82 additions and 2 deletions.
diff --git a/CHANGELOG.rst b/CHANGELOG.rst
@@ -7,6 +7,14 @@
 Changelog
 =========
 
+0.9.0 (2024-07-xx)
+------------------
+
+**New features**
+
+* Added :meth:`metalearners.metalearner.MetaLearner.init_params`.
+
+
 0.8.0 (2024-07-22)
 ------------------
 

diff --git a/metalearners/drlearner.py b/metalearners/drlearner.py
@@ -4,7 +4,7 @@
 
 import numpy as np
 from joblib import Parallel, delayed
-from typing_extensions import Self
+from typing_extensions import Any, Self
 
 from metalearners._typing import (
     Features,
@@ -398,3 +398,7 @@ def _pseudo_outcome(
             )
 
         return pseudo_outcome
+
+    @property
+    def init_args(self) -> dict[str, Any]:
+        return super().init_args | {"adaptive_clipping": self.adaptive_clipping}
diff --git a/metalearners/metalearner.py b/metalearners/metalearner.py
@@ -5,7 +5,7 @@
 from collections.abc import Callable, Collection, Sequence
 from copy import deepcopy
 from dataclasses import dataclass
-from typing import TypedDict
+from typing import Any, TypedDict
 
 import numpy as np
 import pandas as pd
@@ -1123,6 +1123,55 @@ def _default_scoring() -> Scoring:
             return default_scoring
         return dict(default_scoring) | dict(scoring)
 
+    @property
+    def init_args(self) -> dict[str, Any]:
+        """Create initiliazation parameters for a new MetaLearner.
+
+        Importantly, this does not copy further internal state, such as the weights or
+        parameters of trained base models.
+        """
+        return {
+            "is_classification": self.is_classification,
+            "n_variants": self.n_variants,
+            "nuisance_model_factory": {
+                k: v
+                for k, v in self.nuisance_model_factory.items()
+                if k != PROPENSITY_MODEL
+                if k not in self._prefitted_nuisance_models
+            },
+            "treatment_model_factory": self.treatment_model_factory,
+            "propensity_model_factory": (
+                self.nuisance_model_factory.get(PROPENSITY_MODEL)
+                if PROPENSITY_MODEL not in self._prefitted_nuisance_models
+                else None
+            ),
+            "nuisance_model_params": {
+                k: v
+                for k, v in self.nuisance_model_params.items()
+                if k != PROPENSITY_MODEL
+                if k not in self._prefitted_nuisance_models
+            },
+            "treatment_model_params": self.treatment_model_params,
+            "propensity_model_params": (
+                self.nuisance_model_params.get(PROPENSITY_MODEL)
+                if PROPENSITY_MODEL not in self._prefitted_nuisance_models
+                else None
+            ),
+            "fitted_nuisance_models": {
+                k: deepcopy(v)
+                for k, v in self._nuisance_models.items()
+                if k in self._prefitted_nuisance_models and k != PROPENSITY_MODEL
+            },
+            "fitted_propensity_model": (
+                deepcopy(self._nuisance_models.get(PROPENSITY_MODEL))
+                if PROPENSITY_MODEL in self._prefitted_nuisance_models
+                else None
+            ),
+            "feature_set": self.feature_set,
+            "n_folds": self.n_folds,
+            "random_state": self.random_state,
+        }
+
 
 class _ConditionalAverageOutcomeMetaLearner(MetaLearner, ABC):
 

diff --git a/tests/test_metalearner.py b/tests/test_metalearner.py
@@ -1119,3 +1119,22 @@ def test_validate_outcome_different_classes(implementation, use_pandas, rng):
         ValueError, match="have seen different sets of classification outcomes."
     ):
         ml.fit(X, y, w)
+
+
+@pytest.mark.parametrize(
+    "implementation",
+    [TLearner, SLearner, XLearner, RLearner, DRLearner],
+)
+def test_init_args(implementation):
+    ml = implementation(
+        True,
+        2,
+        LogisticRegression,
+        LinearRegression,
+        LogisticRegression,
+    )
+    ml2 = implementation(**ml.init_args)
+
+    assert set(ml.__dict__.keys()) == set(ml2.__dict__.keys())
+    for key in ml.__dict__:
+        assert ml.__dict__[key] == ml2.__dict__[key]