clean up docstrings

Author: Pranav Choudhary

pyaptamer/aptanet/_feature_classifier.py

@@ -50,29 +50,25 @@ class AptaNetClassifier(ClassifierMixin, BaseEstimator):
     eps : float, default=1e-08
         Epsilon value for numerical stability in the optimizer.
     weight_decay : float, default=0.0
-        L2 regularisation (weight decay) applied by the optimizer. Increasing
-        this value helps prevent the neural network from overfitting.
+        L2 regularization penalty applied by the optimizer.
     n_estimators : int, default=300
-        Number of trees in the ``RandomForestClassifier`` used for feature
-        selection. Ignored when a custom ``estimator`` is provided.
+        Number of trees in the `RandomForestClassifier` used for feature
+        selection. Ignored when a custom `estimator` is provided.
     max_depth : int or None, default=9
-        Maximum depth of each tree in the ``RandomForestClassifier`` used for
-        feature selection. Ignored when a custom ``estimator`` is provided.
+        Maximum depth of each tree in the `RandomForestClassifier` used for
+        feature selection. Ignored when a custom `estimator` is provided.
     optimizer : torch.optim.Optimizer class, default=torch.optim.RMSprop
-        PyTorch optimizer class to use for training the neural network.
+        Optimizer class passed to skorch for neural network training.
     device : str or None, default=None
-        Device string passed to skorch (e.g. ``"cpu"``, ``"cuda"``,
-        ``"cuda:1"``). When ``None``, the device is selected automatically:
-        ``"cuda"`` if a GPU is available, otherwise ``"cpu"``.
+        Device string for skorch (e.g. `"cpu"`, `"cuda"`, `"cuda:1"`).
+        If `None`, uses `"cuda"` when a GPU is available, else `"cpu"`.
     estimator : sklearn estimator or None, default=None
-        Estimator used for feature selection via ``SelectFromModel``. When
-        ``None``, a ``RandomForestClassifier`` with ``n_estimators`` and
-        ``max_depth`` is used.
+        Estimator passed to `SelectFromModel`. If `None`, uses a
+        `RandomForestClassifier` with `n_estimators` and `max_depth`.
     random_state : int or None, default=None
-        Random seed for reproducibility. When set, both NumPy and Torch seeds
-        are fixed.
+        Seed for NumPy and Torch random generators.
     threshold : str or float, default="mean"
-        Threshold passed to ``SelectFromModel`` (e.g. ``"mean"`` or a float).
+        Feature-selection threshold passed to `SelectFromModel`.
     verbose : int, default=0
         Verbosity level for the underlying skorch neural net.
 
@@ -280,29 +276,25 @@ class AptaNetRegressor(RegressorMixin, BaseEstimator):
     eps : float, default=1e-08
         Epsilon value for numerical stability in the optimizer.
     weight_decay : float, default=0.0
-        L2 regularisation (weight decay) applied by the optimizer. Increasing
-        this value helps prevent the neural network from overfitting.
+        L2 regularization penalty applied by the optimizer.
     n_estimators : int, default=300
-        Number of trees in the ``RandomForestRegressor`` used for feature
-        selection. Ignored when a custom ``estimator`` is provided.
+        Number of trees in the `RandomForestRegressor` used for feature
+        selection. Ignored when a custom `estimator` is provided.
     max_depth : int or None, default=9
-        Maximum depth of each tree in the ``RandomForestRegressor`` used for
-        feature selection. Ignored when a custom ``estimator`` is provided.
+        Maximum depth of each tree in the `RandomForestRegressor` used for
+        feature selection. Ignored when a custom `estimator` is provided.
     optimizer : torch.optim.Optimizer class, default=torch.optim.RMSprop
-        PyTorch optimizer class to use for training the neural network.
+        Optimizer class passed to skorch for neural network training.
     device : str or None, default=None
-        Device string passed to skorch (e.g. ``"cpu"``, ``"cuda"``,
-        ``"cuda:1"``). When ``None``, the device is selected automatically:
-        ``"cuda"`` if a GPU is available, otherwise ``"cpu"``.
+        Device string for skorch (e.g. `"cpu"`, `"cuda"`, `"cuda:1"`).
+        If `None`, uses `"cuda"` when a GPU is available, else `"cpu"`.
     estimator : sklearn estimator or None, default=None
-        Estimator used for feature selection via ``SelectFromModel``. When
-        ``None``, a ``RandomForestRegressor`` with ``n_estimators`` and
-        ``max_depth`` is used.
+        Estimator passed to `SelectFromModel`. If `None`, uses a
+        `RandomForestRegressor` with `n_estimators` and `max_depth`.
     random_state : int or None, default=None
-        Random seed for reproducibility. When set, both NumPy and Torch seeds
-        are fixed.
+        Seed for NumPy and Torch random generators.
     threshold : str or float, default="mean"
-        Threshold passed to ``SelectFromModel`` (e.g. ``"mean"`` or a float).
+        Feature-selection threshold passed to `SelectFromModel`.
     verbose : int, default=0
         Verbosity level for the underlying skorch neural net.
     """

pyaptamer/benchmarking/_base.py

@@ -47,16 +47,13 @@ class Benchmarking:
             - "test"  = mean of cross_validate(...)[f"test_{metric}"]
 
     raw_results_ : pd.DataFrame or None
-        Per-fold scores produced by :meth:`run` when ``return_raw=True``.
+        Per-fold scores, populated after every :meth:`run` call.
 
         - Index: pandas.MultiIndex with three levels
             - level 0 "estimator": estimator name
             - level 1 "metric": evaluator name
             - level 2 "fold": fold index (0-based)
         - Columns: ["train", "test"] (both floats)
-        - Cell values: raw per-fold scores, directly compatible with
-          ``sktime``'s ``Evaluator`` for Friedman tests and Critical
-          Difference diagrams.
 
     Example
     -------
@@ -123,11 +120,7 @@ def _to_df(self, results):
         return pd.DataFrame(records, index=index, columns=["train", "test"])
 
     def _to_raw_df(self, raw_results):
-        """Convert nested per-fold results to a raw DataFrame.
-
-        The resulting DataFrame is directly compatible with ``sktime``'s
-        ``Evaluator`` class for Friedman tests and Critical Difference diagrams.
-        """
+        """Convert nested per-fold results to a raw DataFrame."""
         records = []
         index = []
 
@@ -151,14 +144,10 @@ def run(self, return_raw=False):
         Parameters
         ----------
         return_raw : bool, default=False
-            If ``False`` (default), returns only a summary DataFrame with
-            mean scores across folds.
-
-            If ``True``, returns a tuple ``(summary, raw)`` where ``raw`` is
-            a per-fold DataFrame with a three-level MultiIndex
-            ``(estimator, metric, fold)``. The ``raw`` DataFrame is directly
-            compatible with ``sktime``'s ``Evaluator`` class for Friedman
-            tests and Critical Difference diagrams.
+            If `False` (default), returns only the summary DataFrame.
+            If `True`, also returns `raw_results_` as the second element
+            of a tuple, containing per-fold scores keyed by
+            ``(estimator, metric, fold)``.
 
         Returns
         -------