perf(qav3): reduce optuna train search space

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

diff --git a/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py b/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
index 6f1fbc30b175f9e7334c967d8ab9ff10071e0a93..388152453e06f4256d08c325f3dfe4a9d370d9da 100644 (file)
--- a/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
+++ b/quickadapter/user_data/freqaimodels/QuickAdapterRegressorV3.py
@@ -1029,21 +1029,11 @@ def train_objective(
     candles_step: int,
     model_training_parameters: dict,
 ) -> float:
-    min_train_window: int = fit_live_predictions_candles * int(round(1 / test_size - 1))
-    max_train_window: int = len(X)
-    if max_train_window < min_train_window:
-        min_train_window = max_train_window
-    train_window: int = trial.suggest_int(
-        "train_period_candles", min_train_window, max_train_window, step=candles_step
-    )
-    X = X.iloc[-train_window:]
-    y = y.iloc[-train_window:]
-    train_weights = train_weights[-train_window:]
-
-    min_test_window: int = fit_live_predictions_candles
+    min_test_window: int = fit_live_predictions_candles * 2
+    if len(X_test) < min_test_window:
+        logger.warning(f"Insufficient test data: {len(X_test)} < {min_test_window}")
+        return np.inf
     max_test_window: int = len(X_test)
-    if max_test_window < min_test_window:
-        min_test_window = max_test_window
     test_window: int = trial.suggest_int(
         "test_period_candles", min_test_window, max_test_window, step=candles_step
     )
@@ -1051,6 +1041,18 @@ def train_objective(
     y_test = y_test.iloc[-test_window:]
     test_weights = test_weights[-test_window:]
 
+    min_train_window: int = min_test_window * int(round(1 / test_size - 1))
+    if len(X) < min_train_window:
+        logger.warning(f"Insufficient train data: {len(X)} < {min_train_window}")
+        return np.inf
+    max_train_window: int = len(X)
+    train_window: int = trial.suggest_int(
+        "train_period_candles", min_train_window, max_train_window, step=candles_step
+    )
+    X = X.iloc[-train_window:]
+    y = y.iloc[-train_window:]
+    train_weights = train_weights[-train_window:]
+
     model = fit_regressor(
         regressor=regressor,
         X=X,
@@ -1063,12 +1065,10 @@ def train_objective(
     )
     y_pred = model.predict(X_test)
 
-    error = sklearn.metrics.root_mean_squared_error(
+    return sklearn.metrics.root_mean_squared_error(
         y_test, y_pred, sample_weight=test_weights
     )
 
-    return error
-
 
 def get_optuna_study_model_parameters(
     trial: optuna.trial.Trial, regressor: str
@@ -1129,12 +1129,10 @@ def hp_objective(
     )
     y_pred = model.predict(X_test)
 
-    error = sklearn.metrics.root_mean_squared_error(
+    return sklearn.metrics.root_mean_squared_error(
         y_test, y_pred, sample_weight=test_weights
     )
 
-    return error
-
 
 def calculate_quantile(values: np.ndarray, value: float) -> float:
     if values.size == 0: