wavetrainer 0.0.42__tar.gz → 0.0.44__tar.gz

{wavetrainer-0.0.42/wavetrainer.egg-info → wavetrainer-0.0.44}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: wavetrainer
-Version: 0.0.42
+Version: 0.0.44
 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-0.0.42 → wavetrainer-0.0.44}/setup.py

@@ -23,7 +23,7 @@ def install_requires() -> typing.List[str]:
 
 setup(
     name='wavetrainer',
-    version='0.0.42',
+    version='0.0.44',
     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-0.0.42 → wavetrainer-0.0.44}/tests/trainer_test.py

@@ -40,7 +40,7 @@ class TestTrainer(unittest.TestCase):
 
     def test_trainer_dt_column(self):
         with tempfile.TemporaryDirectory() as tmpdir:
-            trainer = Trainer(tmpdir, walkforward_timedelta=datetime.timedelta(days=7), trials=1, dt_column="dt_column")
+            trainer = Trainer(tmpdir, walkforward_timedelta=datetime.timedelta(days=7), trials=5, dt_column="dt_column")
             x_data = [i for i in range(100)]
             x_index = [datetime.datetime(2022, 1, 1) + datetime.timedelta(days=i) for i in range(len(x_data))]
             df = pd.DataFrame(

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/__init__.py

@@ -2,5 +2,5 @@
 
 from .create import create
 
-__VERSION__ = "0.0.42"
+__VERSION__ = "0.0.44"
 __all__ = ("create",)

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/calibrator/calibrator_router.py

@@ -37,6 +37,9 @@ class CalibratorRouter(Calibrator):
         return "router"
 
     def predictions_as_x(self, y: pd.Series | pd.DataFrame | None = None) -> bool:
+        calibrator = self._calibrator
+        if calibrator is not None:
+            return calibrator.predictions_as_x(None)
         if y is None:
             raise ValueError("y is null")
         if determine_model_type(y) == ModelType.REGRESSION:

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/calibrator/mapie_calibrator.py

@@ -21,7 +21,7 @@ class MAPIECalibrator(Calibrator):
 
     def __init__(self, model: Model):
         super().__init__(model)
-        self._mapie = MapieRegressor(model.create_estimator(), method="plus")
+        self._mapie = MapieRegressor(model.create_estimator(), method="plus", cv=5)
 
     @classmethod
     def name(cls) -> str:
@@ -54,19 +54,22 @@ class MAPIECalibrator(Calibrator):
             raise ValueError("mapie is null")
         if y is None:
             raise ValueError("y is null")
-        if len(df) <= 5:
-            return self
-        mapie.fit(df.to_numpy(), y.to_numpy())
+        mapie.fit(self._model.convert_df(df), y)
         return self
 
     def transform(self, df: pd.DataFrame) -> pd.DataFrame:
         alpha = []
         for potential_alpha in [0.05, 0.32]:
-            if len(df) > int(1.0 / potential_alpha) + 1:
+            if (
+                len(df) > int(1.0 / potential_alpha) + 1
+                and len(df) > int(1.0 / (1.0 - potential_alpha)) + 1
+            ):
                 alpha.append(potential_alpha)
         ret_df = pd.DataFrame(index=df.index)
         if alpha:
-            _, y_pis = self._mapie.predict(df, alpha=alpha)
+            _, y_pis = self._mapie.predict(
+                self._model.convert_df(df), alpha=alpha, allow_infinite_bounds=True
+            )
             for i in range(y_pis.shape[1]):
                 if i >= len(alpha):
                     continue

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/model/catboost/catboost_kwargs.py

@@ -3,6 +3,7 @@
 from typing import Any
 
 import numpy as np
+import pandas as pd
 from catboost import Pool  # type: ignore
 
 ORIGINAL_X_ARG_KEY = "original_x"
@@ -12,12 +13,17 @@ CAT_FEATURES_ARG_KEY = "cat_features"
 
 def handle_fit_kwargs(*args, **kwargs) -> tuple[tuple[Any, ...], dict[str, Any]]:
     """Handles keyword args coming into a catboost fit method."""
+    args_list = list(args)
+    fit_x = args_list[0]
+
+    cat_features = kwargs.get(CAT_FEATURES_ARG_KEY)
+    if cat_features is None and isinstance(fit_x, pd.DataFrame):
+        cat_features = fit_x.select_dtypes(include="category").columns.tolist()
+    kwargs[CAT_FEATURES_ARG_KEY] = cat_features
+
     if ORIGINAL_X_ARG_KEY in kwargs:
         df = kwargs[ORIGINAL_X_ARG_KEY]
         eval_x, eval_y = kwargs[EVAL_SET_ARG_KEY]
-        cat_features = kwargs[CAT_FEATURES_ARG_KEY]
-        args_list = list(args)
-        fit_x = args_list[0]
         fix_x_cp = fit_x.copy()
 
         # Stupid code to ensure eval is feature equivalent to train data
@@ -32,9 +38,6 @@ def handle_fit_kwargs(*args, **kwargs) -> tuple[tuple[Any, ...], dict[str, Any]]
                     included_columns.append(col)
                     break
         # We also need to update cat_features or catboost will yell at us
-        cat_features = list(
-            set(list(kwargs.get(CAT_FEATURES_ARG_KEY, []))) & set(included_columns)
-        )
         args_list[0] = df[included_columns]
         args = tuple(args_list)
 
@@ -45,7 +48,7 @@ def handle_fit_kwargs(*args, **kwargs) -> tuple[tuple[Any, ...], dict[str, Any]]
                 label=eval_y,
                 cat_features=cat_features,
             )
-        kwargs[CAT_FEATURES_ARG_KEY] = cat_features
 
         del kwargs[ORIGINAL_X_ARG_KEY]
+
     return args, kwargs

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/model/catboost/catboost_model.py

@@ -2,12 +2,12 @@
 
 # pylint: disable=line-too-long
 import json
-import logging
 import os
 from typing import Self
 
 import optuna
 import pandas as pd
+import pytest_is_running
 import torch
 from catboost import CatBoost, Pool  # type: ignore
 
@@ -83,12 +83,21 @@ class CatboostModel(Model):
     def create_estimator(self):
         return self._create_catboost()
 
+    def reset(self):
+        self._catboost = None
+        self._best_iteration = None
+
+    def convert_df(self, df: pd.DataFrame) -> pd.DataFrame:
+        return df
+
     def set_options(
         self, trial: optuna.Trial | optuna.trial.FrozenTrial, df: pd.DataFrame
     ) -> None:
         self._iterations = trial.suggest_int(_ITERATIONS_KEY, 100, 10000)
         self._learning_rate = trial.suggest_float(_LEARNING_RATE_KEY, 0.001, 0.3)
-        self._depth = trial.suggest_int(_DEPTH_KEY, 1, 6)
+        self._depth = trial.suggest_int(
+            _DEPTH_KEY, 1, 2 if pytest_is_running.is_running() else 6
+        )
         self._l2_leaf_reg = trial.suggest_float(_L2_LEAF_REG_KEY, 3.0, 50.0)
         self._boosting_type = trial.suggest_categorical(
             _BOOSTING_TYPE_KEY, ["Ordered", "Plain"]
@@ -170,8 +179,6 @@ class CatboostModel(Model):
             metric_period=100,
             eval_set=eval_pool,
         )
-        importances = catboost.get_feature_importance(prettified=True)
-        logging.info("Importances:\n%s", importances)
         self._best_iteration = catboost.get_best_iteration()
         return self
 

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/model/model.py

@@ -46,3 +46,11 @@ class Model(Params, Fit):
     def create_estimator(self) -> Any:
         """Creates a new estimator."""
         raise NotImplementedError("creates_estimator not implemented in parent class.")
+
+    def reset(self) -> None:
+        """Resets a model."""
+        raise NotImplementedError("reset not implemented in parent class.")
+
+    def convert_df(self, df: pd.DataFrame) -> pd.DataFrame:
+        """Converts a dataframe for use with a model."""
+        raise NotImplementedError("convert_df not implemented in parent class.")

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/model/model_router.py

@@ -66,6 +66,18 @@ class ModelRouter(Model):
             raise ValueError("model is null")
         return model.create_estimator()
 
+    def reset(self):
+        model = self._model
+        if model is None:
+            raise ValueError("model is null")
+        model.reset()
+
+    def convert_df(self, df: pd.DataFrame) -> pd.DataFrame:
+        model = self._model
+        if model is None:
+            raise ValueError("model is null")
+        return model.convert_df(df)
+
     def set_options(
         self, trial: optuna.Trial | optuna.trial.FrozenTrial, df: pd.DataFrame
     ) -> None:

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/model/tabpfn/tabpfn_model.py

@@ -56,6 +56,12 @@ class TabPFNModel(Model):
     def create_estimator(self):
         return self._create_tabpfn()
 
+    def reset(self):
+        pass
+
+    def convert_df(self, df: pd.DataFrame) -> pd.DataFrame:
+        return df
+
     def set_options(
         self, trial: optuna.Trial | optuna.trial.FrozenTrial, df: pd.DataFrame
     ) -> None:

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/model/xgboost/xgboost_model.py

@@ -7,6 +7,7 @@ from typing import Self
 
 import optuna
 import pandas as pd
+import pytest_is_running
 import torch
 from xgboost import XGBClassifier, XGBRegressor
 from xgboost.callback import TrainingCallback
@@ -107,7 +108,11 @@ class XGBoostModel(Model):
     @property
     def feature_importances(self) -> dict[str, float]:
         bst = self._provide_xgboost()
-        return bst.get_booster().get_score(importance_type="weight")  # type: ignore
+        try:
+            return bst.get_booster().get_score(importance_type="weight")  # type: ignore
+        except XGBoostError as exc:
+            print(str(exc))
+            return {}
 
     def provide_estimator(self):
         return self._provide_xgboost()
@@ -115,6 +120,13 @@ class XGBoostModel(Model):
     def create_estimator(self):
         return self._create_xgboost()
 
+    def reset(self):
+        self._xgboost = None
+        self._best_iteration = None
+
+    def convert_df(self, df: pd.DataFrame) -> pd.DataFrame:
+        return _convert_categoricals(df)
+
     def set_options(
         self, trial: optuna.Trial | optuna.trial.FrozenTrial, df: pd.DataFrame
     ) -> None:
@@ -126,7 +138,9 @@ class XGBoostModel(Model):
         self._subsample = trial.suggest_float(_SUBSAMPLE_KEY, 0.2, 1.0)
         self._colsample_bytree = trial.suggest_float(_COLSAMPLE_BYTREE_KEY, 0.2, 1.0)
         if self._booster in ["gbtree", "dart"]:
-            self._max_depth = trial.suggest_int(_MAX_DEPTH_KEY, 3, 9)
+            self._max_depth = trial.suggest_int(
+                _MAX_DEPTH_KEY, 3, 4 if pytest_is_running.is_running() else 9
+            )
             self._min_child_weight = trial.suggest_int(
                 _MIN_CHILD_WEIGHT_KEY, 2, 10, log=True
             )
@@ -144,7 +158,9 @@ class XGBoostModel(Model):
             )
             self._rate_drop = trial.suggest_float(_RATE_DROP_KEY, 1e-8, 1.0, log=True)
             self._skip_drop = trial.suggest_float(_SKIP_DROP_KEY, 1e-8, 1.0, log=True)
-        self._num_boost_rounds = trial.suggest_int(_NUM_BOOST_ROUNDS_KEY, 100, 10000)
+        self._num_boost_rounds = trial.suggest_int(
+            _NUM_BOOST_ROUNDS_KEY, 100, 110 if pytest_is_running.is_running() else 10000
+        )
         self._early_stopping_rounds = trial.suggest_int(
             _EARLY_STOPPING_ROUNDS_KEY, 50, 500
         )
@@ -275,7 +291,8 @@ class XGBoostModel(Model):
             )
         param = {
             "objective": "binary:logistic",
-            "tree_method": "gpu_hist" if torch.cuda.is_available() else "exact",
+            "tree_method": "hist" if torch.cuda.is_available() else "exact",
+            "device": "cuda" if torch.cuda.is_available() else "cpu",
             # defines booster, gblinear for linear functions.
             "booster": self._booster,
             # L2 regularization weight.

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/reducer/base_selector_reducer.py

@@ -7,7 +7,6 @@ import joblib  # type: ignore
 import optuna
 import pandas as pd
 from feature_engine.selection.base_selector import BaseSelector
-from sklearn.utils.validation import check_is_fitted  # type: ignore
 
 from .reducer import Reducer
 
@@ -60,6 +59,4 @@ class BaseSelectorReducer(Reducer):
     def transform(self, df: pd.DataFrame) -> pd.DataFrame:
         if len(df.columns) <= 1:
             return df
-        if not check_is_fitted(self._base_selector):
-            return df
         return self._base_selector.transform(df)

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/reducer/combined_reducer.py

@@ -1,5 +1,6 @@
 """A reducer that combines all the other reducers."""
 
+# pylint: disable=line-too-long
 import json
 import logging
 import os
@@ -107,8 +108,8 @@ class CombinedReducer(Reducer):
             removed_columns = before_columns.difference(after_columns)
             if removed_columns:
                 removed_columns_dict[reducer.name()] = list(removed_columns)
-            logging.info(
-                "%s reducer took %f", reducer.name(), time.time() - start_reducer
+            print(
+                f"{reducer.name()} reducer took {time.time() - start_reducer} and removed {len(removed_columns)} features",
             )
         if self._folder is not None:
             with open(

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/reducer/correlation_reducer.py

@@ -37,7 +37,7 @@ class CorrelationReducer(BaseSelectorReducer):
         self, trial: optuna.Trial | optuna.trial.FrozenTrial, df: pd.DataFrame
     ) -> None:
         self._correlation_selector.threshold = trial.suggest_float(
-            _CORRELATION_REDUCER_THRESHOLD, 0.1, 0.9
+            _CORRELATION_REDUCER_THRESHOLD, 0.7, 0.99
         )
 
     def fit(

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/reducer/smart_correlation_reducer.py

@@ -33,7 +33,7 @@ class SmartCorrelationReducer(BaseSelectorReducer):
         self, trial: optuna.Trial | optuna.trial.FrozenTrial, df: pd.DataFrame
     ) -> None:
         self._correlation_selector.threshold = trial.suggest_float(
-            _SMART_CORRELATION_REDUCER_THRESHOLD, 0.1, 0.9
+            _SMART_CORRELATION_REDUCER_THRESHOLD, 0.7, 0.99
         )
 
     def fit(
@@ -48,3 +48,8 @@ class SmartCorrelationReducer(BaseSelectorReducer):
         if len(self._correlation_selector.variables) <= 1:
             return self
         return super().fit(df, y=y, w=w, eval_x=eval_x, eval_y=eval_y)
+
+    def transform(self, df: pd.DataFrame) -> pd.DataFrame:
+        if len(find_non_categorical_numeric_columns(df)) <= 1:
+            return df
+        return super().transform(df)

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/selector/selector.py

@@ -57,7 +57,8 @@ class Selector(Params, Fit):
             return self
         if not isinstance(y, pd.Series):
             raise ValueError("y is not a series.")
-        if len(df.columns) <= 1:
+        total_columns = len(df.columns)
+        if total_columns <= 1:
             return self
         print(
             f"Performing feature selection with {self._steps} steps and a total ratio of {self._feature_ratio}"
@@ -81,6 +82,9 @@ class Selector(Params, Fit):
             if not current_features:
                 current_features = [list(feature_importances.keys())[0]]
             current_features = current_features[:required_features]
+            print(
+                f"Current Features:\n{pd.Series(data=list(feature_importances.values()), index=list(feature_importances.keys()))}\n"
+            )
 
         n_features = len(current_features)
         for i in range(self._steps):
@@ -90,16 +94,18 @@ class Selector(Params, Fit):
             ratio_diff = 1.0 - self._feature_ratio
             ratio_step = ratio_diff / float(self._steps)
             current_ratio = 1.0 - (ratio_step * i)
-            n_features = max(1, int(len(df.columns) * current_ratio))
+            n_features = max(1, int(total_columns * current_ratio))
             if n_features >= len(current_features):
                 continue
 
+            self._model.reset()
             self._model.fit(df, y=y, w=w, eval_x=eval_x, eval_y=eval_y)
             set_current_features(n_features)
             print(f"Reduced features to {len(current_features)}")
             df = df[current_features]
             if eval_x is not None:
                 eval_x = eval_x[current_features]
+        print(f"Final feature count: {len(current_features)}")
 
         self._selector = current_features
 

{wavetrainer-0.0.42 → wavetrainer-0.0.44}/wavetrainer/trainer.py

@@ -212,7 +212,7 @@ class Trainer(Fit):
                 folder = os.path.join(
                     self._folder, str(y_series.name), split_idx.isoformat()
                 )
-                new_folder = os.path.exists(folder)
+                new_folder = not os.path.exists(folder)
                 os.makedirs(folder, exist_ok=True)
                 trial_file = os.path.join(folder, _TRIAL_FILENAME)
                 if os.path.exists(trial_file):
@@ -281,13 +281,14 @@ class Trainer(Fit):
                     x_test = selector.transform(x_test)
                     print(f"Selection took {time.time() - start_selector}")
                     start_train = time.time()
-                    x_pred = model.fit_transform(
+                    model.fit(
                         x_train,
                         y=y_train,
                         w=w,
                         eval_x=x_test if not no_evaluation else None,
                         eval_y=y_test if not no_evaluation else None,
                     )
+                    y_pred = model.transform(x_test)
                     print(f"Training took {time.time() - start_train}")
 
                     # Calibrate
@@ -295,13 +296,12 @@ class Trainer(Fit):
                     calibrator = CalibratorRouter(model)
                     calibrator.set_options(trial, x)
                     calibrator.fit(
-                        x_pred if calibrator.predictions_as_x(y_train) else x_train,
-                        y=y_train,
+                        y_pred if calibrator.predictions_as_x(y_test) else x_test,
+                        y=y_test,
                     )
                     print(f"Calibrating took {time.time() - start_calibrate}")
 
                     # Output
-                    y_pred = model.transform(x_test)
                     cal_pred = calibrator.transform(
                         y_pred if calibrator.predictions_as_x(y_test) else x_test
                     )
@@ -441,6 +441,8 @@ class Trainer(Fit):
                         if self._max_train_timeout is None
                         else self._max_train_timeout.total_seconds(),
                     )
+                else:
+                    break
 
                 _fit(study.best_trial, test_df, test_series, True, test_idx, True)
                 last_processed_dt = test_idx
@@ -515,7 +517,9 @@ class Trainer(Fit):
                 x_pred = reducer.transform(group[feature_columns])
                 x_pred = selector.transform(x_pred)
                 y_pred = model.transform(x_pred)
-                y_pred = calibrator.transform(y_pred)
+                y_pred = calibrator.transform(
+                    y_pred if calibrator.predictions_as_x(None) else x_pred
+                )
                 for new_column in y_pred.columns.values:
                     group["_".join([column, new_column])] = y_pred[new_column]
                 return group

{wavetrainer-0.0.42 → wavetrainer-0.0.44/wavetrainer.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: wavetrainer
-Version: 0.0.42
+Version: 0.0.44
 Summary: A library for automatically finding the optimal model within feature and hyperparameter space.
 Home-page: https://github.com/8W9aG/wavetrainer
 Author: Will Sackfield