Fix moo things

autosklearn/automl.py

@@ -1502,6 +1502,7 @@ def fit_ensemble(
         ensemble_nbest: Optional[int] = None,
         ensemble_class: Optional[AbstractEnsemble] = EnsembleSelection,
         ensemble_kwargs: Optional[Dict[str, Any]] = None,
+        metrics: Scorer | Sequence[Scorer] | None = None,
     ):
         check_is_fitted(self)
 
@@ -1532,6 +1533,10 @@ def fit_ensemble(
         else:
             self._is_dask_client_internally_created = False
 
+        metrics = metrics if metrics is not None else self._metrics
+        if not isinstance(metrics, Sequence):
+            metrics = [metrics]
+
         # Use the current thread to start the ensemble builder process
         # The function ensemble_builder_process will internally create a ensemble
         # builder in the provide dask client
@@ -1541,7 +1546,7 @@ def fit_ensemble(
             backend=copy.deepcopy(self._backend),
             dataset_name=dataset_name if dataset_name else self._dataset_name,
             task=task if task else self._task,
-            metrics=self._metrics,
+            metrics=metrics if metrics is not None else self._metrics,
             ensemble_class=(
                 ensemble_class if ensemble_class is not None else self._ensemble_class
             ),
@@ -1652,20 +1657,14 @@ def _load_best_individual_model(self):
         return ensemble
 
     def _load_pareto_set(self) -> Sequence[VotingClassifier | VotingRegressor]:
-        if len(self._metrics) <= 1:
-            raise ValueError("Pareto set is only available for two or more metrics.")
-
         if self._ensemble_class is not None:
             self.ensemble_ = self._backend.load_ensemble(self._seed)
         else:
             self.ensemble_ = None
 
         # If no ensemble is loaded we cannot do anything
         if not self.ensemble_:
-
-            raise ValueError(
-                "Pareto set can only be accessed if an ensemble is available."
-            )
+            raise ValueError("Pareto set only available if ensemble can be loaded.")
 
         if isinstance(self.ensemble_, AbstractMultiObjectiveEnsemble):
             pareto_set = self.ensemble_.get_pareto_set()

autosklearn/ensemble_building/builder.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 
-from typing import Any, Dict, Iterable, Sequence, Type, cast
+from typing import Any, Iterable, Mapping, Sequence, Type, cast
 
 import logging.handlers
 import multiprocessing
@@ -46,7 +46,7 @@ def __init__(
         task_type: int,
         metrics: Sequence[Scorer],
         ensemble_class: Type[AbstractEnsemble] = EnsembleSelection,
-        ensemble_kwargs: Dict[str, Any] | None = None,
+        ensemble_kwargs: Mapping[str, Any] | None = None,
         ensemble_nbest: int | float = 50,
         max_models_on_disc: int | float | None = 100,
         seed: int = 1,
@@ -71,9 +71,11 @@ def __init__(
         metrics: Sequence[Scorer]
             Metrics to optimize the ensemble for. These must be non-duplicated.
 
-        ensemble_class
+        ensemble_class: Type[AbstractEnsemble]
+            Implementation of the ensemble algorithm.
 
-        ensemble_kwargs
+        ensemble_kwargs: Mapping[str, Any] | None
+            Arguments passed to the constructor of the ensemble algorithm.
 
         ensemble_nbest: int | float = 50
 
@@ -169,6 +171,8 @@ def __init__(
         self.validation_performance_ = np.inf
 
         # Data we may need
+        # TODO: The test data is needlessly loaded but automl_common has no concept of
+        # these and is perhaps too rigid
         datamanager: XYDataManager = self.backend.load_datamanager()
         self._X_test: SUPPORTED_FEAT_TYPES | None = datamanager.data.get("X_test", None)
         self._y_test: np.ndarray | None = datamanager.data.get("Y_test", None)
@@ -442,6 +446,17 @@ def main(
             self.logger.debug("Found no runs")
             raise RuntimeError("Found no runs")
 
+        # We load in `X_data` if we need it
+        if any(m._needs_X for m in self.metrics):
+            ensemble_X_data = self.X_data("ensemble")
+
+            if ensemble_X_data is None:
+                msg = "No `X_data` for 'ensemble' which was required by metrics"
+                self.logger.debug(msg)
+                raise RuntimeError(msg)
+        else:
+            ensemble_X_data = None
+
         # Calculate the loss for those that require it
         requires_update = self.requires_loss_update(runs)
         if self.read_at_most is not None:
@@ -450,9 +465,7 @@ def main(
         for run in requires_update:
             run.record_modified_times()  # So we don't count as modified next time
             run.losses = {
-                metric.name: self.loss(
-                    run, metric=metric, X_data=self.X_data("ensemble")
-                )
+                metric.name: self.loss(run, metric=metric, X_data=ensemble_X_data)
                 for metric in self.metrics
             }
 
@@ -549,15 +562,14 @@ def main(
             return self.ensemble_history, self.ensemble_nbest
 
         targets = cast(np.ndarray, self.targets("ensemble"))  # Sure they exist
-        X_data = self.X_data("ensemble")
 
         ensemble = self.fit_ensemble(
             candidates=candidates,
-            X_data=X_data,
             targets=targets,
             runs=runs,
             ensemble_class=self.ensemble_class,
             ensemble_kwargs=self.ensemble_kwargs,
+            X_data=ensemble_X_data,
             task=self.task_type,
             metrics=self.metrics,
             precision=self.precision,
@@ -587,7 +599,15 @@ def main(
 
             run_preds = [r.predictions(kind, precision=self.precision) for r in models]
             pred = ensemble.predict(run_preds)
-            X_data = self.X_data(kind)
+
+            if any(m._needs_X for m in self.metrics):
+                X_data = self.X_data(kind)
+                if X_data is None:
+                    msg = f"No `X` data for '{kind}' which was required by metrics"
+                    self.logger.debug(msg)
+                    raise RuntimeError(msg)
+            else:
+                X_data = None
 
             scores = calculate_scores(
                 solution=pred_targets,
@@ -597,10 +617,19 @@ def main(
                 X_data=X_data,
                 scoring_functions=None,
             )
+
+            # TODO only one metric in history
+            #
+            #   We should probably return for all metrics but this makes
+            #   automl::performance_history a lot more complicated, will
+            #   tackle in a future PR
+            first_metric = self.metrics[0]
             performance_stamp[f"ensemble_{score_name}_score"] = scores[
-                self.metrics[0].name
+                first_metric.name
             ]
-            self.ensemble_history.append(performance_stamp)
+
+        # Add the performance stamp to the history
+        self.ensemble_history.append(performance_stamp)
 
         # Lastly, delete any runs that need to be deleted. We save this as the last step
         # so that we have an ensemble saved that is up to date. If we do not do so,
@@ -805,13 +834,13 @@ def candidate_selection(
 
     def fit_ensemble(
         self,
-        candidates: list[Run],
-        X_data: SUPPORTED_FEAT_TYPES,
-        targets: np.ndarray,
+        candidates: Sequence[Run],
+        runs: Sequence[Run],
         *,
-        runs: list[Run],
+        targets: np.ndarray | None = None,
         ensemble_class: Type[AbstractEnsemble] = EnsembleSelection,
-        ensemble_kwargs: Dict[str, Any] | None = None,
+        ensemble_kwargs: Mapping[str, Any] | None = None,
+        X_data: SUPPORTED_FEAT_TYPES | None = None,
         task: int | None = None,
         metrics: Sequence[Scorer] | None = None,
         precision: int | None = None,
@@ -825,24 +854,24 @@ def fit_ensemble(
 
         Parameters
         ----------
-        candidates: list[Run]
+        candidates: Sequence[Run]
             List of runs to build an ensemble from
 
-        X_data: SUPPORTED_FEAT_TYPES
-            The base level data.
+        runs: Sequence[Run]
+            List of all runs (also pruned ones and dummy runs)
 
-        targets: np.ndarray
+        targets: np.ndarray | None = None
             The targets to build the ensemble with
 
-        runs: list[Run]
-            List of all runs (also pruned ones and dummy runs)
-
-        ensemble_class: AbstractEnsemble
+        ensemble_class: Type[AbstractEnsemble]
             Implementation of the ensemble algorithm.
 
-        ensemble_kwargs: Dict[str, Any]
+        ensemble_kwargs: Mapping[str, Any] | None
             Arguments passed to the constructor of the ensemble algorithm.
 
+        X_data: SUPPORTED_FEAT_TYPES | None = None
+            The base level data.
+
         task: int | None = None
             The kind of task performed
 
@@ -859,24 +888,42 @@ def fit_ensemble(
         -------
         AbstractEnsemble
         """
-        task = task if task is not None else self.task_type
+        # Validate we have targets if None specified
+        if targets is None:
+            targets = self.targets("ensemble")
+            if targets is None:
+                path = self.backend._get_targets_ensemble_filename()
+                raise ValueError(f"`fit_ensemble` could not find any targets at {path}")
+
         ensemble_class = (
             ensemble_class if ensemble_class is not None else self.ensemble_class
         )
-        ensemble_kwargs = (
-            ensemble_kwargs if ensemble_kwargs is not None else self.ensemble_kwargs
-        )
-        ensemble_kwargs = ensemble_kwargs if ensemble_kwargs is not None else {}
-        metrics = metrics if metrics is not None else self.metrics
-        rs = random_state if random_state is not None else self.random_state
 
-        ensemble = ensemble_class(
-            task_type=task,
-            metrics=metrics,
-            random_state=rs,
-            backend=self.backend,
-            **ensemble_kwargs,
-        )  # type: AbstractEnsemble
+        # Create the ensemble_kwargs, favouring in order:
+        # 1) function kwargs, 2) function params 3) init_kwargs 4) init_params
+
+        # Collect func params in dict if they're not None
+        params = {
+            k: v
+            for k, v in [
+                ("task_type", task),
+                ("metrics", metrics),
+                ("random_state", random_state),
+            ]
+            if v is not None
+        }
+
+        kwargs = {
+            "backend": self.backend,
+            "task_type": self.task_type,
+            "metrics": self.metrics,
+            "random_state": self.random_state,
+            **(self.ensemble_kwargs or {}),
+            **params,
+            **(ensemble_kwargs or {}),
+        }
+
+        ensemble = ensemble_class(**kwargs)  # type: AbstractEnsemble
 
         self.logger.debug(f"Fitting ensemble on {len(candidates)} models")
         start_time = time.time()
@@ -995,7 +1042,8 @@ def loss(
         self,
         run: Run,
         metric: Scorer,
-        X_data: SUPPORTED_FEAT_TYPES,
+        *,
+        X_data: SUPPORTED_FEAT_TYPES | None = None,
         kind: str = "ensemble",
     ) -> float:
         """Calculate the loss for a run
@@ -1008,6 +1056,9 @@ def loss(
         metric: Scorer
             The metric to calculate the loss of
 
+        X_data: SUPPORTED_FEAT_TYPES | None = None
+            Any X_data required to be passed to the metric
+
         kind: str = "ensemble"
             The kind of targets to use for the run
 

autosklearn/estimators.py

@@ -601,6 +601,7 @@ def fit_ensemble(
         ensemble_kwargs: Optional[Dict[str, Any]] = None,
         ensemble_nbest: Optional[int] = None,
         ensemble_class: Optional[AbstractEnsemble] = EnsembleSelection,
+        metrics: Scorer | Sequence[Scorer] | None = None,
     ):
         """Fit an ensemble to models trained during an optimization process.
 
@@ -650,12 +651,13 @@ def fit_ensemble(
             to obtain only use the single best model instead of an
             ensemble.
 
+        metrics: Scorer | Sequence[Scorer] | None = None
+            A metric or list of metrics to score the ensemble with
+
         Returns
         -------
         self
-
         """
-
         # User specified `ensemble_size` explicitly, warn them about deprecation
         if ensemble_size is not None:
             # Keep consistent behaviour
@@ -708,6 +710,7 @@ def fit_ensemble(
             ensemble_nbest=ensemble_nbest,
             ensemble_class=ensemble_class,
             ensemble_kwargs=ensemble_kwargs,
+            metrics=metrics,
         )
         return self
 

test/fixtures/ensemble_building.py

@@ -164,6 +164,10 @@ def _make(
                 backend.save_additional_data(
                     datamanager.data["Y_train"], what="targets_ensemble"
                 )
+            if "X_train" in datamanager.data:
+                backend.save_additional_data(
+                    datamanager.data["X_train"], what="input_ensemble"
+                )
 
         builder = EnsembleBuilder(
             backend=backend,

test/fixtures/metrics.py

@@ -0,0 +1,26 @@
+from typing import Any
+
+import numpy as np
+
+from autosklearn.metrics import accuracy, make_scorer
+
+
+def _accuracy_requiring_X_data(
+    y_true: np.ndarray,
+    y_pred: np.ndarray,
+    X_data: Any,
+) -> float:
+    """Dummy metric that needs X Data"""
+    if X_data is None:
+        raise ValueError()
+    return accuracy(y_true, y_pred)
+
+
+acc_with_X_data = make_scorer(
+    name="acc_with_X_data",
+    score_func=_accuracy_requiring_X_data,
+    needs_X=True,
+    optimum=1,
+    worst_possible_result=0,
+    greater_is_better=True,
+)

test/test_automl/__init__.py

@@ -1 +0,0 @@
-# -*- encoding: utf-8 -*-

test/test_automl/test_construction.py

@@ -1,9 +1,4 @@
-"""Property based Tests
-
-These test are for checking properties of already fitted models. Any test that does
-tests using cases should not modify the state as these models are cached between tests
-to reduce training time.
-"""
+"""Test things related to only constructing an AutoML instance"""
 from typing import Any, Dict, Optional, Union
 
 from autosklearn.automl import AutoML

test/test_automl/test_dataset_compression.py

@@ -1,3 +1,4 @@
+"""Test things related to how AutoML compresses the dataset size"""
 from typing import Any, Callable, Dict
 
 import numpy as np