workbench 0.8.201__py3-none-any.whl → 0.8.204__py3-none-any.whl

workbench/model_scripts/pytorch_model/pytorch.template

@@ -13,17 +13,19 @@ from pytorch_tabular.models import CategoryEmbeddingModelConfig
 # Model Performance Scores
 from sklearn.metrics import (
     mean_absolute_error,
+    median_absolute_error,
     r2_score,
     root_mean_squared_error,
     precision_recall_fscore_support,
     confusion_matrix,
 )
+from scipy.stats import spearmanr
 
 # Classification Encoder
 from sklearn.preprocessing import LabelEncoder
 
 # Scikit Learn Imports
-from sklearn.model_selection import train_test_split
+from sklearn.model_selection import train_test_split, KFold, StratifiedKFold
 
 from io import StringIO
 import json
@@ -36,9 +38,9 @@ TEMPLATE_PARAMS = {
     "model_type": "{{model_type}}",
     "target": "{{target_column}}",
     "features": "{{feature_list}}",
+    "id_column": "{{id_column}}",
     "compressed_features": "{{compressed_features}}",
     "model_metrics_s3_path": "{{model_metrics_s3_path}}",
-    "train_all_data": "{{train_all_data}}",
     "hyperparameters": "{{hyperparameters}}",
 }
 
@@ -204,36 +206,57 @@ def decompress_features(
     return df, decompressed_features
 
 
-def model_fn(model_dir: str) -> TabularModel:
-    """Load the PyTorch Tabular model from the specified directory.
+def model_fn(model_dir: str) -> dict:
+    """Load the PyTorch Tabular ensemble models from the specified directory.
 
     Args:
-        model_dir: Directory containing the saved model
+        model_dir: Directory containing the saved model(s)
 
     Returns:
-        Loaded TabularModel instance
+        Dictionary with ensemble models and metadata
     """
+    import torch
+    from functools import partial
+
+    # Load ensemble metadata if present
+    ensemble_metadata_path = os.path.join(model_dir, "ensemble_metadata.joblib")
+    if os.path.exists(ensemble_metadata_path):
+        ensemble_metadata = joblib.load(ensemble_metadata_path)
+        n_ensemble = ensemble_metadata["n_ensemble"]
+    else:
+        n_ensemble = 1
+
+    # Determine map_location for loading models (handle CUDA trained models on CPU inference)
+    map_location = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    # Patch torch.load globally to use map_location (needed for joblib-loaded callbacks)
+    # This handles the case where pytorch-tabular loads callbacks.sav via joblib,
+    # which internally calls torch.load without map_location
+    original_torch_load = torch.load
+    torch.load = partial(original_torch_load, map_location=map_location)
+
     # Save current working directory
     original_cwd = os.getcwd()
+    ensemble_models = []
+
     try:
         # Change to /tmp because Pytorch Tabular needs write access (creates a .pt_tmp directory)
         os.chdir("/tmp")
 
-        # Remove callbacks.sav if it exists - it's not needed for inference and causes
-        # GPU->CPU loading issues (joblib.load doesn't support map_location)
-        model_path = os.path.join(model_dir, "tabular_model")
-        callbacks_path = os.path.join(model_path, "callbacks.sav")
-        if os.path.exists(callbacks_path):
-            os.remove(callbacks_path)
-
-        # Load the model (map_location="cpu" ensures GPU-trained models work on CPU endpoints)
-        model = TabularModel.load_model(model_path, map_location="cpu")
+        for ens_idx in range(n_ensemble):
+            # Try numbered model path first, fall back to legacy path
+            model_path = os.path.join(model_dir, f"tabular_model_{ens_idx}")
+            if not os.path.exists(model_path):
+                model_path = os.path.join(model_dir, "tabular_model")
+            model = TabularModel.load_model(model_path, map_location=map_location)
+            ensemble_models.append(model)
 
     finally:
-        # Restore the original working directory
+        # Restore torch.load and working directory
+        torch.load = original_torch_load
         os.chdir(original_cwd)
 
-    return model
+    return {"ensemble_models": ensemble_models, "n_ensemble": n_ensemble}
 
 
 def input_fn(input_data, content_type: str) -> pd.DataFrame:
@@ -264,18 +287,23 @@ def output_fn(output_df: pd.DataFrame, accept_type: str) -> tuple[str, str]:
         raise RuntimeError(f"{accept_type} accept type is not supported by this script.")
 
 
-def predict_fn(df: pd.DataFrame, model: TabularModel) -> pd.DataFrame:
-    """Make Predictions with our PyTorch Tabular Model
+def predict_fn(df: pd.DataFrame, model_dict: dict) -> pd.DataFrame:
+    """Make Predictions with our PyTorch Tabular Model ensemble.
 
     Args:
         df (pd.DataFrame): The input DataFrame
-        model: The TabularModel use for predictions
+        model_dict: Dictionary containing ensemble models and metadata
 
     Returns:
-        pd.DataFrame: The DataFrame with the predictions added
+        pd.DataFrame: The DataFrame with predictions (and prediction_std for ensembles)
     """
+    model_type = TEMPLATE_PARAMS["model_type"]
     compressed_features = TEMPLATE_PARAMS["compressed_features"]
 
+    # Extract ensemble models
+    ensemble_models = model_dict["ensemble_models"]
+    n_ensemble = model_dict["n_ensemble"]
+
     # Grab our feature columns (from training)
     model_dir = os.environ.get("SM_MODEL_DIR", "/opt/ml/model")
     with open(os.path.join(model_dir, "feature_columns.json")) as fp:
@@ -308,8 +336,10 @@ def predict_fn(df: pd.DataFrame, model: TabularModel) -> pd.DataFrame:
     if missing_mask.any():
         print(f"Warning: {missing_mask.sum()} rows have missing features, will return NaN predictions")
 
-    # Initialize prediction column with NaN
+    # Initialize prediction columns
     df["prediction"] = np.nan
+    if model_type in ["regressor", "uq_regressor"]:
+        df["prediction_std"] = np.nan
 
     # Only predict on complete rows
     complete_df = matched_df[~missing_mask]
@@ -317,37 +347,63 @@ def predict_fn(df: pd.DataFrame, model: TabularModel) -> pd.DataFrame:
         print("Warning: No complete rows to predict on")
         return df
 
-    # Make predictions using the TabularModel
-    result = model.predict(complete_df[features])
-
     # pytorch-tabular returns predictions using f"{target}_prediction" column
     target = TEMPLATE_PARAMS["target"]
     prediction_column = f"{target}_prediction"
-    if prediction_column in result.columns:
-        predictions = result[prediction_column].values
-    else:
-        raise ValueError(f"Cannot find prediction column in: {result.columns.tolist()}")
 
-    # If we have a label encoder, decode the predictions
-    if label_encoder:
-        predictions = label_encoder.inverse_transform(predictions.astype(int))
+    # Collect predictions from all ensemble members
+    all_ensemble_preds = []
+    all_ensemble_probs = []
 
-    # Set predictions only for complete rows
-    df.loc[~missing_mask, "prediction"] = predictions
+    for ens_idx, ens_model in enumerate(ensemble_models):
+        result = ens_model.predict(complete_df[features])
+
+        if prediction_column in result.columns:
+            ens_preds = result[prediction_column].values
+        else:
+            raise ValueError(f"Cannot find prediction column in: {result.columns.tolist()}")
 
-    # For classification, get probabilities
+        all_ensemble_preds.append(ens_preds)
+
+        # For classification, collect probabilities
+        if label_encoder is not None:
+            prob_cols = sorted([col for col in result.columns if col.endswith("_probability")])
+            if prob_cols:
+                all_ensemble_probs.append(result[prob_cols].values)
+
+    # Stack and compute mean/std (std is 0 for single model)
+    ensemble_preds = np.stack(all_ensemble_preds, axis=0)  # (n_ensemble, n_samples)
+    preds = np.mean(ensemble_preds, axis=0)
+    preds_std = np.std(ensemble_preds, axis=0)  # Will be 0s for n_ensemble=1
+
+    print(f"Inference: Ensemble predictions shape: {preds.shape}, n_ensemble: {n_ensemble}")
+
+    # Handle classification vs regression
     if label_encoder is not None:
-        prob_cols = [col for col in result.columns if col.endswith("_probability")]
-        if prob_cols:
-            probs = result[prob_cols].values
+        # For classification, average probabilities then take argmax
+        if all_ensemble_probs:
+            ensemble_probs = np.stack(all_ensemble_probs, axis=0)  # (n_ensemble, n_samples, n_classes)
+            avg_probs = np.mean(ensemble_probs, axis=0)  # (n_samples, n_classes)
+            class_preds = np.argmax(avg_probs, axis=1)
+            predictions = label_encoder.inverse_transform(class_preds)
 
             # Build full proba Series with None for missing rows
             all_proba = pd.Series([None] * len(df), index=df.index, dtype=object)
-            all_proba.loc[~missing_mask] = [p.tolist() for p in probs]
+            all_proba.loc[~missing_mask] = [p.tolist() for p in avg_probs]
             df["pred_proba"] = all_proba
 
             # Expand the pred_proba column into separate columns for each class
             df = expand_proba_column(df, label_encoder.classes_)
+        else:
+            # No probabilities, use averaged predictions
+            predictions = label_encoder.inverse_transform(preds.astype(int))
+    else:
+        # Regression (includes uq_regressor)
+        predictions = preds
+        df.loc[~missing_mask, "prediction_std"] = preds_std
+
+    # Set predictions only for complete rows
+    df.loc[~missing_mask, "prediction"] = predictions
 
     return df
 
@@ -359,12 +415,11 @@ if __name__ == "__main__":
     target = TEMPLATE_PARAMS["target"]
     features = TEMPLATE_PARAMS["features"]
     orig_features = features.copy()
+    id_column = TEMPLATE_PARAMS["id_column"]
     compressed_features = TEMPLATE_PARAMS["compressed_features"]
     model_type = TEMPLATE_PARAMS["model_type"]
     model_metrics_s3_path = TEMPLATE_PARAMS["model_metrics_s3_path"]
-    train_all_data = TEMPLATE_PARAMS["train_all_data"]
     hyperparameters = TEMPLATE_PARAMS["hyperparameters"]
-    validation_split = 0.2
 
     # Script arguments for input/output directories
     parser = argparse.ArgumentParser()
@@ -423,72 +478,71 @@ if __name__ == "__main__":
     # Cast continuous columns to float
     all_df[continuous_cols] = all_df[continuous_cols].astype("float64")
 
-    # Do we want to train on all the data?
-    if train_all_data:
-        print("Training on ALL of the data")
-        df_train = all_df.copy()
-        df_val = all_df.copy()
-
-    # Does the dataframe have a training column?
-    elif "training" in all_df.columns:
-        print("Found training column, splitting data based on training column")
-        df_train = all_df[all_df["training"]].copy()
-        df_val = all_df[~all_df["training"]].copy()
-    else:
-        # Just do a random training Split
-        print("WARNING: No training column found, splitting data with random state=42")
-        df_train, df_val = train_test_split(all_df, test_size=validation_split, random_state=42)
-    print(f"FIT/TRAIN: {df_train.shape}")
-    print(f"VALIDATION: {df_val.shape}")
-
-    # Set up PyTorch Tabular configuration
-    data_config = DataConfig(
-        target=[target],
-        continuous_cols=continuous_cols,
-        categorical_cols=categorical_cols,
-    )
-
-    # Choose the 'task' based on model type also set up the label encoder if needed
+    # Choose the 'task' based on model type and set up the label encoder if needed
     if model_type == "classifier":
         task = "classification"
-        # Encode the target column
+        # Encode the target column on full dataset for consistent encoding
         label_encoder = LabelEncoder()
-        df_train[target] = label_encoder.fit_transform(df_train[target])
-        df_val[target] = label_encoder.transform(df_val[target])
+        all_df[target] = label_encoder.fit_transform(all_df[target])
+        num_classes = len(label_encoder.classes_)
     else:
         task = "regression"
         label_encoder = None
+        num_classes = None
 
     # Use any hyperparameters to set up both the trainer and model configurations
     print(f"Hyperparameters: {hyperparameters}")
+    n_folds = hyperparameters.get("n_folds", 5)  # Number of CV folds (default: 5)
+
+    # =========================================================================
+    # UNIFIED TRAINING: Works for n_folds=1 (single model) or n_folds>1 (K-fold CV)
+    # =========================================================================
+    print(f"Training {'single model' if n_folds == 1 else f'{n_folds}-fold cross-validation ensemble'}...")
+
+    # Create fold splits
+    if n_folds == 1:
+        # Single fold: use train/val split from "training" column or random split
+        if "training" in all_df.columns:
+            print("Found training column, splitting data based on training column")
+            train_idx = np.where(all_df["training"])[0]
+            val_idx = np.where(~all_df["training"])[0]
+        else:
+            print("WARNING: No training column found, splitting data with random 80/20 split")
+            indices = np.arange(len(all_df))
+            train_idx, val_idx = train_test_split(indices, test_size=0.2, random_state=42)
+        folds = [(train_idx, val_idx)]
+    else:
+        # K-Fold CV
+        if model_type == "classifier":
+            kfold = StratifiedKFold(n_splits=n_folds, shuffle=True, random_state=42)
+            split_target = all_df[target]
+        else:
+            kfold = KFold(n_splits=n_folds, shuffle=True, random_state=42)
+            split_target = None
+        folds = list(kfold.split(all_df, split_target))
+
+    # Initialize storage for out-of-fold predictions
+    oof_predictions = np.full(len(all_df), np.nan, dtype=np.float64)
+    if model_type == "classifier" and num_classes and num_classes > 1:
+        oof_proba = np.full((len(all_df), num_classes), np.nan, dtype=np.float64)
+    else:
+        oof_proba = None
 
-    # Set up PyTorch Tabular configuration with defaults
-    trainer_defaults = {
-        "auto_lr_find": False,
-        "batch_size": min(128, max(32, len(df_train) // 16)),
-        "max_epochs": 100,
-        "min_epochs": 10,
-        "early_stopping": "valid_loss",
-        "early_stopping_patience": 10,
-        "checkpoints": "valid_loss",
-        "accelerator": "auto",
-        "progress_bar": "none",
-        "gradient_clip_val": 1.0,
-    }
+    ensemble_models = []
 
-    # Override defaults with training_config if present
-    training_overrides = {k: v for k, v in hyperparameters.get("training_config", {}).items() if k in trainer_defaults}
-    for key, value in training_overrides.items():
-        print(f"TRAINING CONFIG Override: {key}: {trainer_defaults[key]} → {value}")
-    trainer_params = {**trainer_defaults, **training_overrides}
-    trainer_config = TrainerConfig(**trainer_params)
+    # Set up PyTorch Tabular data configuration (shared across folds)
+    data_config = DataConfig(
+        target=[target],
+        continuous_cols=continuous_cols,
+        categorical_cols=categorical_cols,
+    )
 
     # Model config defaults
     model_defaults = {
         "layers": "256-128-64",
         "activation": "LeakyReLU",
         "learning_rate": 1e-3,
-        "dropout": 0.3,
+        "dropout": 0.1,
         "use_batch_norm": True,
         "initialization": "kaiming",
     }
@@ -498,41 +552,107 @@ if __name__ == "__main__":
         print(f"MODEL CONFIG Override: {key}: {model_defaults[key]} → {value}")
     model_params = {**model_defaults, **model_overrides}
 
-    # Use CategoryEmbedding model configuration for general-purpose tabular modeling.
     model_config = CategoryEmbeddingModelConfig(task=task, **model_params)
     optimizer_config = OptimizerConfig()
 
-    #####################################
-    # Create and train the TabularModel #
-    #####################################
-    tabular_model = TabularModel(
-        data_config=data_config,
-        model_config=model_config,
-        optimizer_config=optimizer_config,
-        trainer_config=trainer_config,
-    )
-    tabular_model.fit(train=df_train, validation=df_val)
+    for fold_idx, (train_idx, val_idx) in enumerate(folds):
+        print(f"\n{'='*50}")
+        print(f"Training Fold {fold_idx + 1}/{len(folds)}")
+        print(f"{'='*50}")
+
+        # Split data for this fold
+        df_train = all_df.iloc[train_idx].reset_index(drop=True)
+        df_val = all_df.iloc[val_idx].reset_index(drop=True)
+
+        print(f"Fold {fold_idx + 1} - Train: {len(df_train)}, Val: {len(df_val)}")
+
+        # Set up PyTorch Tabular trainer configuration (per-fold for batch_size)
+        # Calculate batch size that avoids single-sample last batch (batch norm requires >1)
+        batch_size = min(128, max(32, len(df_train) // 16))
+        if len(df_train) % batch_size == 1:
+            batch_size += 1  # Adjust to avoid last batch of size 1
+        trainer_defaults = {
+            "auto_lr_find": False,
+            "batch_size": batch_size,
+            "max_epochs": 200,
+            "min_epochs": 10,
+            "early_stopping": "valid_loss",
+            "early_stopping_patience": 20,
+            "checkpoints": "valid_loss",
+            "accelerator": "auto",
+            "progress_bar": "none",
+            "gradient_clip_val": 1.0,
+            "seed": 42 + fold_idx,
+        }
+
+        # Override defaults with training_config if present
+        training_overrides = {k: v for k, v in hyperparameters.get("training_config", {}).items() if k in trainer_defaults}
+        if fold_idx == 0:  # Only print overrides once
+            for key, value in training_overrides.items():
+                print(f"TRAINING CONFIG Override: {key}: {trainer_defaults[key]} → {value}")
+        trainer_params = {**trainer_defaults, **training_overrides}
+        trainer_config = TrainerConfig(**trainer_params)
+
+        # Create and train the TabularModel for this fold
+        tabular_model = TabularModel(
+            data_config=data_config,
+            model_config=model_config,
+            optimizer_config=optimizer_config,
+            trainer_config=trainer_config,
+        )
+        tabular_model.fit(train=df_train, validation=df_val)
+        ensemble_models.append(tabular_model)
+
+        # Make out-of-fold predictions
+        result = tabular_model.predict(df_val, include_input_features=False)
+        fold_preds = result[f"{target}_prediction"].values
+
+        # Store out-of-fold predictions
+        if model_type == "classifier":
+            oof_predictions[val_idx] = fold_preds.astype(int)
+            prob_cols = sorted([col for col in result.columns if col.endswith("_probability")])
+            if prob_cols and oof_proba is not None:
+                oof_proba[val_idx] = result[prob_cols].values
+        else:
+            oof_predictions[val_idx] = fold_preds.flatten()
 
-    # Make Predictions on the Validation Set
-    print("Making Predictions on Validation Set...")
-    result = tabular_model.predict(df_val, include_input_features=False)
+        print(f"Fold {fold_idx + 1} complete!")
 
-    # pytorch-tabular returns predictions using f"{target}_prediction" column
-    preds = result[f"{target}_prediction"].values
+    print(f"\nTraining complete! Trained {len(ensemble_models)} model(s).")
+
+    # Use out-of-fold predictions for metrics
+    # For n_folds=1, we only have predictions for val_idx, so filter to those rows
+    if n_folds == 1:
+        val_mask = ~np.isnan(oof_predictions)
+        preds = oof_predictions[val_mask]
+        df_val = all_df[val_mask].copy()
+        if oof_proba is not None:
+            oof_proba = oof_proba[val_mask]
+    else:
+        preds = oof_predictions
+        df_val = all_df.copy()
+
+    # Compute prediction_std by running all ensemble models on validation data
+    # For n_folds=1, std will be 0 (only one model). For n_folds>1, std shows ensemble disagreement.
+    preds_std = None
+    if model_type in ["regressor", "uq_regressor"] and len(ensemble_models) > 0:
+        print("Computing prediction_std from ensemble predictions on validation data...")
+        all_ensemble_preds_for_std = []
+        for ens_model in ensemble_models:
+            result = ens_model.predict(df_val[features], include_input_features=False)
+            ens_preds = result[f"{target}_prediction"].values.flatten()
+            all_ensemble_preds_for_std.append(ens_preds)
+
+        ensemble_preds_stacked = np.stack(all_ensemble_preds_for_std, axis=0)
+        preds_std = np.std(ensemble_preds_stacked, axis=0)
+        print(f"Ensemble prediction_std - mean: {np.mean(preds_std):.4f}, max: {np.max(preds_std):.4f}")
 
     if model_type == "classifier":
         # Get probabilities for classification
-        print("Processing Probabilities...")
-        prob_cols = sorted([col for col in result.columns if col.endswith("_probability")])
-        if prob_cols:
-            probs = result[prob_cols].values
-            df_val = df_val.copy()  # Avoid SettingWithCopyWarning
-            df_val["pred_proba"] = [p.tolist() for p in probs]
-
-            # Expand the pred_proba column into separate columns for each class
-            print(df_val.columns.tolist())
+        if oof_proba is not None:
+            df_val = df_val.copy()
+            df_val["pred_proba"] = [p.tolist() for p in oof_proba]
             df_val = expand_proba_column(df_val, label_encoder.classes_)
-            print(df_val.columns.tolist())
 
         # Decode the target and prediction labels
         y_validate = label_encoder.inverse_transform(df_val[target])
@@ -544,7 +664,22 @@ if __name__ == "__main__":
     # Save predictions to S3
     df_val = df_val.copy()
     df_val["prediction"] = preds_decoded
-    output_columns = [target, "prediction"]
+
+    # Build output columns - include id_column if it exists
+    output_columns = []
+    if id_column in df_val.columns:
+        output_columns.append(id_column)
+    output_columns += [target, "prediction"]
+
+    # Add prediction_std for regression models (always present, 0 for single model)
+    if model_type in ["regressor", "uq_regressor"]:
+        if preds_std is not None:
+            df_val["prediction_std"] = preds_std
+        else:
+            df_val["prediction_std"] = 0.0
+        output_columns.append("prediction_std")
+        print(f"Ensemble std - mean: {df_val['prediction_std'].mean():.4f}, max: {df_val['prediction_std'].max():.4f}")
+
     output_columns += [col for col in df_val.columns if col.endswith("_proba")]
     wr.s3.to_csv(
         df_val[output_columns],
@@ -589,14 +724,29 @@ if __name__ == "__main__":
         # Calculate various model performance metrics (regression)
         rmse = root_mean_squared_error(y_validate, preds_decoded)
         mae = mean_absolute_error(y_validate, preds_decoded)
+        medae = median_absolute_error(y_validate, preds_decoded)
         r2 = r2_score(y_validate, preds_decoded)
-        print(f"RMSE: {rmse:.3f}")
-        print(f"MAE: {mae:.3f}")
-        print(f"R2: {r2:.3f}")
-        print(f"NumRows: {len(df_val)}")
+        spearman_corr = spearmanr(y_validate, preds_decoded).correlation
+        support = len(df_val)
+        print(f"rmse: {rmse:.3f}")
+        print(f"mae: {mae:.3f}")
+        print(f"medae: {medae:.3f}")
+        print(f"r2: {r2:.3f}")
+        print(f"spearmanr: {spearman_corr:.3f}")
+        print(f"support: {support}")
+
+    # Save ensemble models
+    for model_idx, ens_model in enumerate(ensemble_models):
+        model_path = os.path.join(args.model_dir, f"tabular_model_{model_idx}")
+        ens_model.save_model(model_path)
+        print(f"Saved model {model_idx + 1} to {model_path}")
+
+    # Save ensemble metadata
+    n_ensemble = len(ensemble_models)
+    ensemble_metadata = {"n_ensemble": n_ensemble, "n_folds": n_folds}
+    joblib.dump(ensemble_metadata, os.path.join(args.model_dir, "ensemble_metadata.joblib"))
+    print(f"Saved ensemble metadata (n_ensemble={n_ensemble}, n_folds={n_folds})")
 
-    # Save the model to the standard place/name
-    tabular_model.save_model(os.path.join(args.model_dir, "tabular_model"))
     if label_encoder:
         joblib.dump(label_encoder, os.path.join(args.model_dir, "label_encoder.joblib"))
 

workbench/model_scripts/uq_models/generated_model_script.py

@@ -5,7 +5,8 @@ from xgboost import XGBRegressor
 from sklearn.model_selection import train_test_split
 
 # Model Performance Scores
-from sklearn.metrics import mean_absolute_error, r2_score, root_mean_squared_error
+from sklearn.metrics import mean_absolute_error, median_absolute_error, r2_score, root_mean_squared_error
+from scipy.stats import spearmanr
 
 from io import StringIO
 import json
@@ -18,11 +19,11 @@ from typing import List, Tuple, Optional, Dict
 
 # Template Placeholders
 TEMPLATE_PARAMS = {
-    "target": "caco_2_efflux",
-    "features": ['chi2v', 'fr_sulfone', 'chi1v', 'bcut2d_logplow', 'fr_piperzine', 'kappa3', 'smr_vsa1', 'slogp_vsa5', 'fr_ketone_topliss', 'fr_sulfonamd', 'fr_imine', 'fr_benzene', 'fr_ester', 'chi2n', 'labuteasa', 'peoe_vsa2', 'smr_vsa6', 'bcut2d_chglo', 'fr_sh', 'peoe_vsa1', 'fr_allylic_oxid', 'chi4n', 'fr_ar_oh', 'fr_nh0', 'fr_term_acetylene', 'slogp_vsa7', 'slogp_vsa4', 'estate_vsa1', 'vsa_estate4', 'numbridgeheadatoms', 'numheterocycles', 'fr_ketone', 'fr_morpholine', 'fr_guanido', 'estate_vsa2', 'numheteroatoms', 'fr_nitro_arom_nonortho', 'fr_piperdine', 'nocount', 'numspiroatoms', 'fr_aniline', 'fr_thiophene', 'slogp_vsa10', 'fr_amide', 'slogp_vsa2', 'fr_epoxide', 'vsa_estate7', 'fr_ar_coo', 'fr_imidazole', 'fr_nitrile', 'fr_oxazole', 'numsaturatedrings', 'fr_pyridine', 'fr_hoccn', 'fr_ndealkylation1', 'numaliphaticheterocycles', 'fr_phenol', 'maxpartialcharge', 'vsa_estate5', 'peoe_vsa13', 'minpartialcharge', 'qed', 'fr_al_oh', 'slogp_vsa11', 'chi0n', 'fr_bicyclic', 'peoe_vsa12', 'fpdensitymorgan1', 'fr_oxime', 'molwt', 'fr_dihydropyridine', 'smr_vsa5', 'peoe_vsa5', 'fr_nitro', 'hallkieralpha', 'heavyatommolwt', 'fr_alkyl_halide', 'peoe_vsa8', 'fr_nhpyrrole', 'fr_isocyan', 'bcut2d_chghi', 'fr_lactam', 'peoe_vsa11', 'smr_vsa9', 'tpsa', 'chi4v', 'slogp_vsa1', 'phi', 'bcut2d_logphi', 'avgipc', 'estate_vsa11', 'fr_coo', 'bcut2d_mwhi', 'numunspecifiedatomstereocenters', 'vsa_estate10', 'estate_vsa8', 'numvalenceelectrons', 'fr_nh2', 'fr_lactone', 'vsa_estate1', 'estate_vsa4', 'numatomstereocenters', 'vsa_estate8', 'fr_para_hydroxylation', 'peoe_vsa3', 'fr_thiazole', 'peoe_vsa10', 'fr_ndealkylation2', 'slogp_vsa12', 'peoe_vsa9', 'maxestateindex', 'fr_quatn', 'smr_vsa7', 'minestateindex', 'numaromaticheterocycles', 'numrotatablebonds', 'fr_ar_nh', 'fr_ether', 'exactmolwt', 'fr_phenol_noorthohbond', 'slogp_vsa3', 'fr_ar_n', 'sps', 'fr_c_o_nocoo', 'bertzct', 'peoe_vsa7', 'slogp_vsa8', 'numradicalelectrons', 'molmr', 'fr_tetrazole', 'numsaturatedcarbocycles', 'bcut2d_mrhi', 'kappa1', 'numamidebonds', 'fpdensitymorgan2', 'smr_vsa8', 'chi1n', 'estate_vsa6', 'fr_barbitur', 'fr_diazo', 'kappa2', 'chi0', 'bcut2d_mrlow', 'balabanj', 'peoe_vsa4', 'numhacceptors', 'fr_sulfide', 'chi3n', 'smr_vsa2', 'fr_al_oh_notert', 'fr_benzodiazepine', 'fr_phos_ester', 'fr_aldehyde', 'fr_coo2', 'estate_vsa5', 'fr_prisulfonamd', 'numaromaticcarbocycles', 'fr_unbrch_alkane', 'fr_urea', 'fr_nitroso', 'smr_vsa10', 'fr_c_s', 'smr_vsa3', 'fr_methoxy', 'maxabspartialcharge', 'slogp_vsa9', 'heavyatomcount', 'fr_azide', 'chi3v', 'smr_vsa4', 'mollogp', 'chi0v', 'fr_aryl_methyl', 'fr_nh1', 'fpdensitymorgan3', 'fr_furan', 'fr_hdrzine', 'fr_arn', 'numaromaticrings', 'vsa_estate3', 'fr_azo', 'fr_halogen', 'estate_vsa9', 'fr_hdrzone', 'numhdonors', 'fr_alkyl_carbamate', 'fr_isothiocyan', 'minabspartialcharge', 'fr_al_coo', 'ringcount', 'chi1', 'estate_vsa7', 'fr_nitro_arom', 'vsa_estate9', 'minabsestateindex', 'maxabsestateindex', 'vsa_estate6', 'estate_vsa10', 'estate_vsa3', 'fr_n_o', 'fr_amidine', 'fr_thiocyan', 'fr_phos_acid', 'fr_c_o', 'fr_imide', 'numaliphaticrings', 'peoe_vsa6', 'vsa_estate2', 'nhohcount', 'numsaturatedheterocycles', 'slogp_vsa6', 'peoe_vsa14', 'fractioncsp3', 'bcut2d_mwlow', 'numaliphaticcarbocycles', 'fr_priamide', 'nacid', 'nbase', 'naromatom', 'narombond', 'sz', 'sm', 'sv', 'sse', 'spe', 'sare', 'sp', 'si', 'mz', 'mm', 'mv', 'mse', 'mpe', 'mare', 'mp', 'mi', 'xch_3d', 'xch_4d', 'xch_5d', 'xch_6d', 'xch_7d', 'xch_3dv', 'xch_4dv', 'xch_5dv', 'xch_6dv', 'xch_7dv', 'xc_3d', 'xc_4d', 'xc_5d', 'xc_6d', 'xc_3dv', 'xc_4dv', 'xc_5dv', 'xc_6dv', 'xpc_4d', 'xpc_5d', 'xpc_6d', 'xpc_4dv', 'xpc_5dv', 'xpc_6dv', 'xp_0d', 'xp_1d', 'xp_2d', 'xp_3d', 'xp_4d', 'xp_5d', 'xp_6d', 'xp_7d', 'axp_0d', 'axp_1d', 'axp_2d', 'axp_3d', 'axp_4d', 'axp_5d', 'axp_6d', 'axp_7d', 'xp_0dv', 'xp_1dv', 'xp_2dv', 'xp_3dv', 'xp_4dv', 'xp_5dv', 'xp_6dv', 'xp_7dv', 'axp_0dv', 'axp_1dv', 'axp_2dv', 'axp_3dv', 'axp_4dv', 'axp_5dv', 'axp_6dv', 'axp_7dv', 'c1sp1', 'c2sp1', 'c1sp2', 'c2sp2', 'c3sp2', 'c1sp3', 'c2sp3', 'c3sp3', 'c4sp3', 'hybratio', 'fcsp3', 'num_stereocenters', 'num_unspecified_stereocenters', 'num_defined_stereocenters', 'num_r_centers', 'num_s_centers', 'num_stereobonds', 'num_e_bonds', 'num_z_bonds', 'stereo_complexity', 'frac_defined_stereo'],
+    "target": "udm_asy_res_efflux_ratio",
+    "features": ['smr_vsa4', 'tpsa', 'nhohcount', 'mollogp', 'peoe_vsa1', 'smr_vsa3', 'nitrogen_span', 'numhdonors', 'minpartialcharge', 'vsa_estate3', 'vsa_estate6', 'tertiary_amine_count', 'hba_hbd_ratio', 'peoe_vsa8', 'estate_vsa4', 'xc_4dv', 'vsa_estate2', 'molmr', 'xp_2dv', 'mi', 'molecular_axis_length', 'vsa_estate4', 'xp_6dv', 'qed', 'estate_vsa8', 'chi1v', 'asphericity', 'axp_1d', 'bcut2d_logphi', 'kappa3', 'axp_7d', 'num_s_centers', 'amphiphilic_moment', 'molecular_asymmetry', 'charge_centroid_distance', 'estate_vsa3', 'vsa_estate8', 'aromatic_interaction_score', 'molecular_volume_3d', 'axp_7dv', 'peoe_vsa3', 'smr_vsa6', 'bcut2d_mrhi', 'radius_of_gyration', 'xpc_4dv', 'minabsestateindex', 'axp_0dv', 'chi4n', 'balabanj', 'bcut2d_mwlow', 'estate_vsa2', 'axp_5d', 'maxestateindex', 'bcut2d_mrlow', 'type_ii_pattern_count', 'avgipc', 'slogp_vsa1', 'fr_nhpyrrole', 'xch_7d', 'axp_1dv', 'peoe_vsa9', 'xch_6d', 'xch_5dv', 'bcut2d_chglo', 'fpdensitymorgan1', 'fr_al_oh', 'axp_5dv', 'smr_vsa5', 'chi2v', 'estate_vsa6', 'smr_vsa9', 'minestateindex', 'bcut2d_logplow', 'c3sp3', 'xp_3d', 'vsa_estate9', 'nbase', 'peoe_vsa2', 'numatomstereocenters', 'xc_5dv', 'bcut2d_mwhi', 'nocount', 'slogp_vsa2', 'smr_vsa1', 'axp_6d', 'maxabspartialcharge', 'vsa_estate5', 'fpdensitymorgan2', 'xp_7d', 'peoe_vsa10', 'num_r_centers', 'mv', 'vsa_estate10', 'xp_0dv', 'axp_4d', 'fractioncsp3', 'smr_vsa10', 'xp_7dv', 'xp_4dv', 'hallkieralpha', 'numhacceptors', 'axp_3d', 'vsa_estate7', 'slogp_vsa3', 'peoe_vsa7', 'estate_vsa10', 'axp_2d', 'c1sp3', 'axp_2dv', 'slogp_vsa4', 'estate_vsa9', 'xch_7dv', 'sps', 'chi0n', 'axp_6dv', 'fr_imidazole', 'xpc_4d', 'bcut2d_chghi', 'chi3n', 'peoe_vsa11', 'xpc_6d', 'estate_vsa1', 'xch_6dv', 'chi3v', 'axp_3dv', 'xc_5d', 'slogp_vsa5', 'maxpartialcharge', 'estate_vsa5', 'fr_hoccn', 'heavyatommolwt', 'fr_ar_n', 'mz', 'xpc_5d', 'axp_4dv', 'xc_3dv', 'xp_6d', 'bertzct', 'peoe_vsa6', 'xc_3d', 'c2sp2', 'fpdensitymorgan3', 'xpc_5dv', 'intramolecular_hbond_potential', 'molwt', 'estate_vsa7', 'xp_5d', 'kappa1', 'xp_5dv', 'chi2n', 'axp_0d', 'xch_4dv', 'xp_4d', 'mp', 'chi1n', 'mm', 'fr_nh0', 'phi', 'labuteasa', 'xp_3dv', 'mse', 'xc_6dv', 'fr_piperzine', 'chi4v', 'xch_4d', 'fr_pyridine', 'xp_2d', 'num_stereocenters', 'minabspartialcharge', 'numaliphaticheterocycles', 'chi0v', 'type_i_pattern_count', 'fr_ketone_topliss', 'fr_ether', 'fr_priamide', 'num_defined_stereocenters', 'mare', 'peoe_vsa12', 'numheteroatoms', 'peoe_vsa4', 'peoe_vsa13', 'hybratio', 'numheterocycles', 'sse', 'fr_c_o_nocoo', 'fr_thiazole', 'slogp_vsa6', 'smr_vsa7', 'vsa_estate1', 'xch_5d', 'smr_vsa2', 'amide_count', 'fr_amide', 'chi1', 'fr_urea', 'fr_aniline', 'fr_aryl_methyl', 'fr_piperdine', 'numrotatablebonds', 'fr_nitrile', 'si', 'sp', 'fr_morpholine', 'numsaturatedheterocycles', 'kappa2', 'chi0', 'mpe', 'sare', 'numaromaticheterocycles', 'naromatom', 'xpc_6dv', 'fr_nh1', 'slogp_vsa10', 'numsaturatedrings', 'c2sp3', 'slogp_vsa8', 'c1sp2', 'fr_ndealkylation2', 'spe', 'xc_6d', 'slogp_vsa11', 'numaliphaticcarbocycles', 'frac_defined_stereo', 'numunspecifiedatomstereocenters', 'fr_benzene', 'xp_1d', 'xc_4d', 'fr_methoxy', 'c4sp3', 'numvalenceelectrons', 'fr_bicyclic', 'fr_imine', 'fr_sulfone', 'c3sp2', 'ringcount', 'slogp_vsa7', 'exactmolwt', 'fr_al_oh_notert', 'fr_tetrazole', 'peoe_vsa14', 'xch_3d', 'fr_para_hydroxylation', 'numspiroatoms'],
     "compressed_features": [],
     "train_all_data": True,
-    "hyperparameters": {},
+    "hyperparameters": {'n_estimators': 500, 'max_depth': 6, 'learning_rate': 0.04},
 }
 
 
@@ -277,10 +278,15 @@ if __name__ == "__main__":
     xgb_mae = mean_absolute_error(y_validate, y_pred_xgb)
     xgb_r2 = r2_score(y_validate, y_pred_xgb)
 
+    xgb_medae = median_absolute_error(y_validate, y_pred_xgb)
+    xgb_spearman = spearmanr(y_validate, y_pred_xgb).correlation
+
     print(f"\nXGBoost Point Prediction Performance:")
-    print(f"RMSE: {xgb_rmse:.3f}")
-    print(f"MAE: {xgb_mae:.3f}")
-    print(f"R2: {xgb_r2:.3f}")
+    print(f"rmse: {xgb_rmse:.3f}")
+    print(f"mae: {xgb_mae:.3f}")
+    print(f"medae: {xgb_medae:.3f}")
+    print(f"r2: {xgb_r2:.3f}")
+    print(f"spearmanr: {xgb_spearman:.3f}")
 
     # Define confidence levels we want to model
     confidence_levels = [0.50, 0.68, 0.80, 0.90, 0.95]  # 50%, 68%, 80%, 90%, 95% confidence intervals
@@ -336,11 +342,14 @@ if __name__ == "__main__":
         coverage = np.mean((y_validate >= y_pis[:, 0, 0]) & (y_validate <= y_pis[:, 1, 0]))
         print(f"  Coverage: Target={confidence_level * 100:.0f}%, Empirical={coverage * 100:.1f}%")
 
+    support = len(df_val)
     print(f"\nOverall Model Performance Summary:")
-    print(f"XGBoost RMSE: {xgb_rmse:.3f}")
-    print(f"XGBoost MAE: {xgb_mae:.3f}")
-    print(f"XGBoost R2: {xgb_r2:.3f}")
-    print(f"NumRows: {len(df_val)}")
+    print(f"rmse: {xgb_rmse:.3f}")
+    print(f"mae: {xgb_mae:.3f}")
+    print(f"medae: {xgb_medae:.3f}")
+    print(f"r2: {xgb_r2:.3f}")
+    print(f"spearmanr: {xgb_spearman:.3f}")
+    print(f"support: {support}")
 
     # Analyze interval widths across confidence levels
     print(f"\nInterval Width Analysis:")