wavetrainer 0.3.18__tar.gz → 0.3.20__tar.gz

{wavetrainer-0.3.18/wavetrainer.egg-info → wavetrainer-0.3.20}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: wavetrainer
-Version: 0.3.18
+Version: 0.3.20
 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-0.3.18 → wavetrainer-0.3.20}/setup.py

@@ -23,7 +23,7 @@ def install_requires() -> typing.List[str]:
 
 setup(
     name='wavetrainer',
-    version='0.3.18',
+    version='0.3.20',
     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-0.3.18 → wavetrainer-0.3.20}/wavetrainer/__init__.py

@@ -2,5 +2,5 @@
 
 from .create import create
 
-__VERSION__ = "0.3.18"
+__VERSION__ = "0.3.20"
 __all__ = ("create",)

wavetrainer-0.3.20/wavetrainer/drift_mixed_sampler.py

@@ -0,0 +1,75 @@
+"""A custom Optuna sampler for dynamic environments."""
+
+from typing import Sequence
+
+import optuna
+
+
+class DriftMixedSampler(optuna.samplers.BaseSampler):
+    """A custom sampler that interleaves TPE with pure random sampling to combat data drift."""
+
+    def __init__(self, random_mix_ratio: float = 0.20, seed: int | None = None):
+        self._tpe_sampler = optuna.samplers.TPESampler(
+            seed=seed, constant_liar=True, multivariate=True
+        )
+        self._random_sampler = optuna.samplers.RandomSampler(seed=seed)
+        self._random_mix_ratio = random_mix_ratio
+
+    def reseed_rng(self) -> None:
+        self._tpe_sampler.reseed_rng()
+        self._random_sampler.reseed_rng()
+
+    def infer_relative_search_space(
+        self, study: optuna.Study, trial: optuna.trial.FrozenTrial
+    ):
+        return self._tpe_sampler.infer_relative_search_space(study, trial)
+
+    def sample_relative(
+        self, study: optuna.Study, trial: optuna.trial.FrozenTrial, search_space
+    ):
+        if (
+            trial.number > 0
+            and self._random_mix_ratio > 0
+            and trial.number % int(1 / self._random_mix_ratio) == 0
+        ):
+            return self._random_sampler.sample_relative(study, trial, search_space)
+        return self._tpe_sampler.sample_relative(study, trial, search_space)
+
+    def sample_independent(
+        self,
+        study: optuna.Study,
+        trial: optuna.trial.FrozenTrial,
+        param_name: str,
+        param_distribution,
+    ):
+        if (
+            trial.number > 0
+            and self._random_mix_ratio > 0
+            and trial.number % int(1 / self._random_mix_ratio) == 0
+        ):
+            return self._random_sampler.sample_independent(
+                study, trial, param_name, param_distribution
+            )
+        return self._tpe_sampler.sample_independent(
+            study, trial, param_name, param_distribution
+        )
+
+    def before_trial(
+        self, study: optuna.Study, trial: optuna.trial.FrozenTrial
+    ) -> None:
+        if hasattr(self._tpe_sampler, "before_trial"):
+            self._tpe_sampler.before_trial(study, trial)
+        if hasattr(self._random_sampler, "before_trial"):
+            self._random_sampler.before_trial(study, trial)
+
+    def after_trial(
+        self,
+        study: optuna.Study,
+        trial: optuna.trial.FrozenTrial,
+        state: optuna.trial.TrialState,
+        values: Sequence[float] | None,
+    ) -> None:
+        if hasattr(self._tpe_sampler, "after_trial"):
+            self._tpe_sampler.after_trial(study, trial, state, values)
+        if hasattr(self._random_sampler, "after_trial"):
+            self._random_sampler.after_trial(study, trial, state, values)

{wavetrainer-0.3.18 → wavetrainer-0.3.20}/wavetrainer/normaliser/powertransformer_normaliser.py

@@ -11,6 +11,7 @@ import numpy as np
 import optuna
 import pandas as pd
 import scipy  # type: ignore
+from sklearn.exceptions import InconsistentVersionWarning  # type: ignore
 from sklearn.preprocessing import PowerTransformer  # type: ignore
 
 from ..exceptions import WavetrainException
@@ -40,7 +41,10 @@ class PowerTransformerNormaliser(Normaliser):
         pass
 
     def load(self, folder: str) -> None:
-        self._pt = joblib.load(os.path.join(folder, _POWERTRANSFORMER_REDUCER_FILE))
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore", category=InconsistentVersionWarning)
+            self._pt = joblib.load(os.path.join(folder, _POWERTRANSFORMER_REDUCER_FILE))
+
         with open(
             os.path.join(folder, _POWERTRANSFORMER_COLUMNS_FILENAME),
             "r",
@@ -88,6 +92,35 @@ class PowerTransformerNormaliser(Normaliser):
                     neginf=0.0,
                 )
         except ValueError as exc:
-            print(exc)
+            # --- Diagnostic Check ---
+            subset = df[self._pt_cols]
+
+            print("\n" + "=" * 50)
+            print("🚨 DATA VALIDATION ERROR DIAGNOSTIC 🚨")
+
+            # Check specifically for infinite values
+            inf_mask = np.isinf(subset)
+            if inf_mask.any().any():
+                bad_cols = [col for col in subset.columns if inf_mask[col].any()]
+                print(f"-> Found infinities (inf or -inf) in columns: {bad_cols}")
+
+                for col in bad_cols:
+                    print(f"\nOffending values in '{col}':")
+                    # Filter and print only the rows where this specific column has an infinity
+                    print(subset[col][inf_mask[col]])
+            else:
+                # Fallback: If it's not strictly 'inf', it's an unusually large number
+                print(
+                    "-> No strict infinities found. Values are likely exceeding float64 capacity."
+                )
+                print(
+                    "\nTop 5 largest absolute values by column to help you hunt it down:"
+                )
+                print(subset.abs().max().sort_values(ascending=False).head(5))
+
+            print("=" * 50 + "\n")
+
+            # Raise the exception as usual so the pipeline stops and you can read the logs
             raise WavetrainException() from exc
+
         return df

{wavetrainer-0.3.18 → wavetrainer-0.3.20}/wavetrainer/trainer.py

@@ -21,6 +21,7 @@ from sklearn.metrics import (accuracy_score, brier_score_loss, log_loss,
 
 from .calibrator.calibrator_router import CalibratorRouter
 from .crps import crps
+from .drift_mixed_sampler import DriftMixedSampler
 from .exceptions import WavetrainException
 from .fit import Fit
 from .model.catboost.catboost_model import CatboostModel
@@ -50,6 +51,7 @@ _IDX_USR_ATTR_KEY = "idx"
 _DT_COLUMN_KEY = "dt_column"
 _MAX_FALSE_POSITIVE_REDUCTION_STEPS_KEY = "max_false_positive_reduction_steps"
 _CORRELATION_CHUNK_SIZE_KEY = "correlation_chunk_size"
+_RANDOM_MIX_RATIO_KEY = "random_mix_ratio"
 _BAD_OUTPUT = -1000.0
 
 
@@ -109,6 +111,7 @@ class Trainer(Fit):
         insert_nulls: bool = False,
         use_power_transformer: bool = False,
         use_correlation_reducer: bool = True,
+        random_mix_ratio: float = 0.20,
         n_jobs: int = 1,
     ):
         tqdm.tqdm.pandas()
@@ -168,6 +171,7 @@ class Trainer(Fit):
                     )
                 if correlation_chunk_size is None:
                     correlation_chunk_size = params.get(_CORRELATION_CHUNK_SIZE_KEY)
+                random_mix_ratio = params.get(_RANDOM_MIX_RATIO_KEY, 0.20)
         else:
             with open(params_file, "w", encoding="utf8") as handle:
                 validation_size_value = None
@@ -200,6 +204,7 @@ class Trainer(Fit):
                         _DT_COLUMN_KEY: dt_column,
                         _MAX_FALSE_POSITIVE_REDUCTION_STEPS_KEY: max_false_positive_reduction_steps,
                         _CORRELATION_CHUNK_SIZE_KEY: correlation_chunk_size,
+                        _RANDOM_MIX_RATIO_KEY: random_mix_ratio,
                     },
                     handle,
                 )
@@ -217,6 +222,7 @@ class Trainer(Fit):
         self._insert_nulls = insert_nulls
         self._use_power_transformer = use_power_transformer
         self._use_correlation_reducer = use_correlation_reducer
+        self._random_mix_ratio = random_mix_ratio
         self._n_jobs = n_jobs
         self._cached_reducers = {}
         self._cached_normalizers = {}
@@ -231,6 +237,12 @@ class Trainer(Fit):
         if os.path.exists(sampler_file):
             with open(sampler_file, "rb") as handle:
                 restored_sampler = pickle.load(handle)
+
+        if restored_sampler is None:
+            restored_sampler = DriftMixedSampler(
+                random_mix_ratio=self._random_mix_ratio
+            )
+
         return optuna.create_study(
             study_name="wavetrain",
             storage=storage_name,
@@ -641,6 +653,10 @@ class Trainer(Fit):
                 last_processed_dt = test_idx
 
             if did_fit:
+                sampler_file = os.path.join(column_dir, _SAMPLER_FILENAME)
+                with open(sampler_file, "wb") as handle:
+                    pickle.dump(study.sampler, handle)
+
                 target_names = ["F1", "Brier"]
                 # fig = optuna.visualization.plot_pareto_front(
                 #    study, target_names=target_names

{wavetrainer-0.3.18 → wavetrainer-0.3.20/wavetrainer.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: wavetrainer
-Version: 0.3.18
+Version: 0.3.20
 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-0.3.18 → wavetrainer-0.3.20}/wavetrainer.egg-info/SOURCES.txt

@@ -10,6 +10,7 @@ tests/model/catboost_kwargs_test.py
 wavetrainer/__init__.py
 wavetrainer/create.py
 wavetrainer/crps.py
+wavetrainer/drift_mixed_sampler.py
 wavetrainer/exceptions.py
 wavetrainer/fit.py
 wavetrainer/model_type.py