refactor(qav3): add a runtime fallback for optuna warm start

Signed-off-by: Jérôme Benoit <jerome.benoit@piment-noir.org>

diff --git a/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py b/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
index 602c8703d29dfced1263504ae6acd4caeb783c91..543856610d9a7805603dd55d39356532bcf80eea 100644 (file)
--- a/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
+++ b/quickadapter/user_data/freqaimodels/LightGBMRegressorQuickAdapterV35.py
@@ -50,6 +50,7 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
             and self.__optuna_config.get("enabled", False)
             and self.data_split_parameters.get("test_size", TEST_SIZE) > 0
         )
+        self.__optuna_hp_params: dict[str, dict] = {}
         self.__optuna_period_params: dict[str, dict] = {}
 
     def fit(self, data_dictionary: Dict, dk: FreqaiDataKitchen, **kwargs) -> Any:
@@ -77,9 +78,12 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
                 dk, X, y, train_weights, X_test, y_test, test_weights
             )
             if optuna_hp_params:
+                if dk.pair not in self.__optuna_hp_params:
+                    self.__optuna_hp_params[dk.pair] = {}
+                self.__optuna_hp_params[dk.pair] = optuna_hp_params
                 model_training_parameters = {
                     **model_training_parameters,
-                    **optuna_hp_params,
+                    **self.__optuna_hp_params[dk.pair],
                 }
 
             optuna_period_params = self.optuna_period_optimize(
@@ -277,7 +281,7 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
         study_name = f"hp-{dk.pair}"
         storage = self.optuna_storage(dk)
         pruner = optuna.pruners.HyperbandPruner()
-        previous_study = self.optuna_previous_study(study_name, storage)
+        previous_study = self.optuna_study_load_and_cleanup(study_name, storage)
         study = optuna.create_study(
             study_name=study_name,
             sampler=optuna.samplers.TPESampler(
@@ -290,6 +294,8 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
         )
         if self.optuna_study_has_best_params(previous_study):
             study.enqueue_trial(previous_study.best_params)
+        elif self.__optuna_hp_params.get(dk.pair):
+            study.enqueue_trial(self.__optuna_hp_params[dk.pair])
         start = time.time()
         try:
             study.optimize(
@@ -334,7 +340,7 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
         study_name = f"period-{dk.pair}"
         storage = self.optuna_storage(dk)
         pruner = optuna.pruners.HyperbandPruner()
-        previous_study = self.optuna_previous_study(study_name, storage)
+        previous_study = self.optuna_study_load_and_cleanup(study_name, storage)
         study = optuna.create_study(
             study_name=study_name,
             sampler=optuna.samplers.TPESampler(
@@ -347,6 +353,10 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
         )
         if self.optuna_study_has_best_params(previous_study):
             study.enqueue_trial(previous_study.best_params)
+        elif self.__optuna_period_params.get(dk.pair):
+            previous_best_params = self.__optuna_period_params[dk.pair].copy()
+            del previous_best_params["rmse"]
+            study.enqueue_trial(previous_best_params)
         start = time.time()
         try:
             study.optimize(
@@ -380,7 +390,7 @@ class LightGBMRegressorQuickAdapterV35(BaseRegressionModel):
             logger.info(f"Optuna period hyperopt | {key:>20s} : {value}")
         return params
 
-    def optuna_previous_study(
+    def optuna_study_load_and_cleanup(
         self, study_name: str, storage
     ) -> optuna.study.Study | None:
         try:

diff --git a/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py b/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
index a43d9156dd1e93773f12d789e1944c405a2dbaf3..7038dffddfcb410cbc85f6c122d317f617ec6dbf 100644 (file)
--- a/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
+++ b/quickadapter/user_data/freqaimodels/XGBoostRegressorQuickAdapterV35.py
@@ -50,6 +50,7 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
             and self.__optuna_config.get("enabled", False)
             and self.data_split_parameters.get("test_size", TEST_SIZE) > 0
         )
+        self.__optuna_hp_params: dict[str, dict] = {}
         self.__optuna_period_params: dict[str, dict] = {}
 
     def fit(self, data_dictionary: Dict, dk: FreqaiDataKitchen, **kwargs) -> Any:
@@ -77,9 +78,12 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
                 dk, X, y, train_weights, X_test, y_test, test_weights
             )
             if optuna_hp_params:
+                if dk.pair not in self.__optuna_hp_params:
+                    self.__optuna_hp_params[dk.pair] = {}
+                self.__optuna_hp_params[dk.pair] = optuna_hp_params
                 model_training_parameters = {
                     **model_training_parameters,
-                    **optuna_hp_params,
+                    **self.__optuna_hp_params[dk.pair],
                 }
 
             optuna_period_params = self.optuna_period_optimize(
@@ -278,7 +282,7 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
         study_name = f"hp-{dk.pair}"
         storage = self.optuna_storage(dk)
         pruner = optuna.pruners.HyperbandPruner()
-        previous_study = self.optuna_previous_study(study_name, storage)
+        previous_study = self.optuna_study_load_and_cleanup(study_name, storage)
         study = optuna.create_study(
             study_name=study_name,
             sampler=optuna.samplers.TPESampler(
@@ -291,6 +295,8 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
         )
         if self.optuna_study_has_best_params(previous_study):
             study.enqueue_trial(previous_study.best_params)
+        elif self.__optuna_hp_params.get(dk.pair):
+            study.enqueue_trial(self.__optuna_hp_params[dk.pair])
         start = time.time()
         try:
             study.optimize(
@@ -335,7 +341,7 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
         study_name = f"period-{dk.pair}"
         storage = self.optuna_storage(dk)
         pruner = optuna.pruners.HyperbandPruner()
-        previous_study = self.optuna_previous_study(study_name, storage)
+        previous_study = self.optuna_study_load_and_cleanup(study_name, storage)
         study = optuna.create_study(
             study_name=study_name,
             sampler=optuna.samplers.TPESampler(
@@ -348,6 +354,10 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
         )
         if self.optuna_study_has_best_params(previous_study):
             study.enqueue_trial(previous_study.best_params)
+        elif self.__optuna_period_params.get(dk.pair):
+            previous_best_params = self.__optuna_period_params[dk.pair].copy()
+            del previous_best_params["rmse"]
+            study.enqueue_trial(previous_best_params)
         start = time.time()
         try:
             study.optimize(
@@ -381,7 +391,7 @@ class XGBoostRegressorQuickAdapterV35(BaseRegressionModel):
             logger.info(f"Optuna period hyperopt | {key:>20s} : {value}")
         return params
 
-    def optuna_previous_study(
+    def optuna_study_load_and_cleanup(
         self, study_name: str, storage
     ) -> optuna.study.Study | None:
         try: