wavetrainer 0.0.29__tar.gz → 0.0.31__tar.gz

{wavetrainer-0.0.29/wavetrainer.egg-info → wavetrainer-0.0.31}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: wavetrainer
-Version: 0.0.29
+Version: 0.0.31
 Summary: A library for automatically finding the optimal model within feature and hyperparameter space.
 Home-page: https://github.com/8W9aG/wavetrainer
 Author: Will Sackfield

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/setup.py

@@ -23,7 +23,7 @@ def install_requires() -> typing.List[str]:
 
 setup(
     name='wavetrainer',
-    version='0.0.29',
+    version='0.0.31',
     description='A library for automatically finding the optimal model within feature and hyperparameter space.',
     long_description=long_description,
     long_description_content_type='text/markdown',

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/tests/trainer_test.py

@@ -58,4 +58,4 @@ class TestTrainer(unittest.TestCase):
             trainer.fit(df, y=y)
             df = trainer.transform(df)
             print("df:")
-            print(df)
+            print(df)

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/__init__.py

@@ -2,5 +2,5 @@
 
 from .create import create
 
-__VERSION__ = "0.0.29"
+__VERSION__ = "0.0.31"
 __all__ = ("create",)

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/model/catboost_model.py

@@ -72,6 +72,16 @@ class CatboostModel(Model):
     def supports_importances(self) -> bool:
         return True
 
+    @property
+    def feature_importances(self) -> dict[str, float]:
+        catboost = self._provide_catboost()
+        importances = catboost.get_feature_importance(prettified=True)
+        if importances is None:
+            raise ValueError("importances is null")
+        feature_ids = importances["Feature Id"].to_list()  # type: ignore
+        importances = importances["Importances"].to_list()  # type: ignore
+        return {feature_ids[x]: importances[x] for x in range(len(feature_ids))}
+
     def pre_fit(
         self,
         df: pd.DataFrame,
@@ -165,7 +175,7 @@ class CatboostModel(Model):
                 label=eval_y,
                 cat_features=eval_x.select_dtypes(include="category").columns.tolist(),
             )
-            if eval_x is not None
+            if eval_x is not None and self._best_iteration is not None
             else None
         )
         catboost.fit(

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/model/model.py

@@ -37,6 +37,13 @@ class Model(Params, Fit):
             "supports_importances not implemented in parent class."
         )
 
+    @property
+    def feature_importances(self) -> dict[str, float]:
+        """The feature importances of this model."""
+        raise NotImplementedError(
+            "feature_importances not implemented in parent class."
+        )
+
     def pre_fit(
         self,
         df: pd.DataFrame,

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/model/model_router.py

@@ -52,6 +52,13 @@ class ModelRouter(Model):
             raise ValueError("model is null")
         return model.supports_importances
 
+    @property
+    def feature_importances(self) -> dict[str, float]:
+        model = self._model
+        if model is None:
+            raise ValueError("model is null")
+        return model.feature_importances
+
     def pre_fit(
         self,
         df: pd.DataFrame,

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/model/tabpfn_model.py

@@ -48,6 +48,10 @@ class TabPFNModel(Model):
     def supports_importances(self) -> bool:
         return False
 
+    @property
+    def feature_importances(self) -> dict[str, float]:
+        return {}
+
     def pre_fit(
         self,
         df: pd.DataFrame,

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/reducer/combined_reducer.py

@@ -90,7 +90,16 @@ class CombinedReducer(Reducer):
         eval_y: pd.Series | pd.DataFrame | None = None,
     ) -> Self:
         for reducer in self._reducers:
+            before_columns = set(df.columns.values)
             df = reducer.fit_transform(df)
+            after_columns = set(df.columns.values)
+            removed_columns = before_columns.difference(after_columns)
+            if removed_columns:
+                logging.info(
+                    "Removed columns %s using %s",
+                    ",".join(removed_columns),
+                    reducer.name(),
+                )
         return self
 
     def transform(self, df: pd.DataFrame) -> pd.DataFrame:

{wavetrainer-0.0.29 → wavetrainer-0.0.31}/wavetrainer/trainer.py

@@ -37,6 +37,14 @@ _TEST_SIZE_KEY = "test_size"
 _VALIDATION_SIZE_KEY = "validation_size"
 _IDX_USR_ATTR_KEY = "idx"
 _DT_COLUMN_KEY = "dt_column"
+_BAD_OUTPUT = -1.0
+
+
+def _assign_bin(timestamp, bins: list[datetime.datetime]) -> int:
+    for i in range(len(bins) - 1):
+        if bins[i] <= timestamp < bins[i + 1]:
+            return i
+    return len(bins) - 2  # Assign to last bin if at the end
 
 
 class Trainer(Fit):
@@ -202,6 +210,7 @@ class Trainer(Fit):
                 folder = os.path.join(
                     self._folder, str(y_series.name), split_idx.isoformat()
                 )
+                new_folder = os.path.exists(folder)
                 os.makedirs(folder, exist_ok=True)
                 trial_file = os.path.join(folder, _TRIAL_FILENAME)
                 if os.path.exists(trial_file):
@@ -227,8 +236,10 @@ class Trainer(Fit):
                     x_train = windower.fit_transform(x_train)
                     y_train = y_train[-len(x_train) :]
                     if len(y_train.unique()) <= 1:
+                        if new_folder:
+                            os.removedirs(folder)
                         logging.warning("Y train only contains 1 unique datapoint.")
-                        return -1.0
+                        return _BAD_OUTPUT
 
                     # Perform common reductions
                     reducer = CombinedReducer()
@@ -290,7 +301,9 @@ class Trainer(Fit):
                     return output
                 except WavetrainException as exc:
                     logging.warning(str(exc))
-                    return -1.0
+                    if new_folder:
+                        os.removedirs(folder)
+                    return _BAD_OUTPUT
 
             start_validation_index = (
                 dt_index.to_list()[-int(len(dt_index) * self._validation_size) - 1]
@@ -332,6 +345,15 @@ class Trainer(Fit):
                     if self._max_train_timeout is None
                     else self._max_train_timeout.total_seconds(),
                 )
+            while study.best_trial.value is None or study.best_trial.value != _BAD_OUTPUT:
+                study.optimize(
+                    test_objective,
+                    n_trials=1,
+                    show_progress_bar=True,
+                    timeout=None
+                    if self._max_train_timeout is None
+                    else self._max_train_timeout.total_seconds(),
+                )
 
             train_len = len(df[dt_index < start_test_index])
             test_len = len(
@@ -426,12 +448,6 @@ class Trainer(Fit):
                 + [(dt_index.max() + pd.Timedelta(days=1)).to_pydatetime()]
             )
 
-            def assign_bin(timestamp, bins: list[datetime.datetime]) -> int:
-                for i in range(len(bins) - 1):
-                    if bins[i] <= timestamp < bins[i + 1]:
-                        return i
-                return len(bins) - 2  # Assign to last bin if at the end
-
             def perform_predictions(
                 group: pd.DataFrame,
                 column_path: str,
@@ -471,7 +487,7 @@ class Trainer(Fit):
 
             old_index = dt_index.copy()
             df = df.groupby(
-                dt_index.map(functools.partial(assign_bin, bins=bins))
+                dt_index.map(functools.partial(_assign_bin, bins=bins))
             ).progress_apply(  # type: ignore
                 functools.partial(
                     perform_predictions,
@@ -489,3 +505,21 @@ class Trainer(Fit):
             df[col] = input_df[col]
 
         return df
+
+    def feature_importances(self) -> dict[str, dict[str, float]]:
+        """Find the feature importances for the rolling models."""
+        feature_importances = {}
+
+        for column in os.listdir(self._folder):
+            column_path = os.path.join(self._folder, column)
+            if not os.path.isdir(column_path):
+                continue
+            for date_str in os.listdir(column_path):
+                date_path = os.path.join(column_path, date_str)
+                if not os.path.isdir(date_path):
+                    continue
+                model = ModelRouter()
+                model.load(date_path)
+                feature_importances[date_str] = model.feature_importances
+
+        return feature_importances

{wavetrainer-0.0.29 → wavetrainer-0.0.31/wavetrainer.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: wavetrainer
-Version: 0.0.29
+Version: 0.0.31
 Summary: A library for automatically finding the optimal model within feature and hyperparameter space.
 Home-page: https://github.com/8W9aG/wavetrainer
 Author: Will Sackfield