workbench 0.8.170__py3-none-any.whl → 0.8.172__py3-none-any.whl

workbench/model_scripts/custom_models/uq_models/generated_model_script.py

@@ -1,8 +1,8 @@
-# Model: NGBoost Regressor with Distribution output
-from ngboost import NGBRegressor
-from xgboost import XGBRegressor  # Base Estimator
+# Model: XGBoost for point predictions + LightGBM with MAPIE for conformalized intervals
+from mapie.regression import ConformalizedQuantileRegressor
+from lightgbm import LGBMRegressor
+from xgboost import XGBRegressor
 from sklearn.model_selection import train_test_split
-import numpy as np
 
 # Model Performance Scores
 from sklearn.metrics import (
@@ -16,20 +16,16 @@ import json
 import argparse
 import joblib
 import os
+import numpy as np
 import pandas as pd
-
-# Local Imports
-from proximity import Proximity
-
-
+from typing import List, Tuple
 
 # Template Placeholders
 TEMPLATE_PARAMS = {
-    "id_column": "id",
-    "features": ['molwt', 'mollogp', 'molmr', 'heavyatomcount', 'numhacceptors', 'numhdonors', 'numheteroatoms', 'numrotatablebonds', 'numvalenceelectrons', 'numaromaticrings', 'numsaturatedrings', 'numaliphaticrings', 'ringcount', 'tpsa', 'labuteasa', 'balabanj', 'bertzct'],
-    "target": "solubility",
-    "train_all_data": True,
-    "track_columns": ['solubility']
+    "target": "udm_asy_res_value",
+    "features": ['bcut2d_logplow', 'numradicalelectrons', 'smr_vsa5', 'fr_lactam', 'fr_morpholine', 'fr_aldehyde', 'slogp_vsa1', 'fr_amidine', 'bpol', 'fr_ester', 'fr_azo', 'kappa3', 'peoe_vsa5', 'fr_ketone_topliss', 'vsa_estate9', 'estate_vsa9', 'bcut2d_mrhi', 'fr_ndealkylation1', 'numrotatablebonds', 'minestateindex', 'fr_quatn', 'peoe_vsa3', 'fr_epoxide', 'fr_aniline', 'minpartialcharge', 'fr_nitroso', 'fpdensitymorgan2', 'fr_oxime', 'fr_sulfone', 'smr_vsa1', 'kappa1', 'fr_pyridine', 'numaromaticrings', 'vsa_estate6', 'molmr', 'estate_vsa1', 'fr_dihydropyridine', 'vsa_estate10', 'fr_alkyl_halide', 'chi2n', 'fr_thiocyan', 'fpdensitymorgan1', 'fr_unbrch_alkane', 'slogp_vsa9', 'chi4n', 'fr_nitro_arom', 'fr_al_oh', 'fr_furan', 'fr_c_s', 'peoe_vsa8', 'peoe_vsa14', 'numheteroatoms', 'fr_ndealkylation2', 'maxabspartialcharge', 'vsa_estate2', 'peoe_vsa7', 'apol', 'numhacceptors', 'fr_tetrazole', 'vsa_estate1', 'peoe_vsa9', 'naromatom', 'bcut2d_chghi', 'fr_sh', 'fr_halogen', 'slogp_vsa4', 'fr_benzodiazepine', 'molwt', 'fr_isocyan', 'fr_prisulfonamd', 'maxabsestateindex', 'minabsestateindex', 'peoe_vsa11', 'slogp_vsa12', 'estate_vsa5', 'numaliphaticcarbocycles', 'bcut2d_mwlow', 'slogp_vsa7', 'fr_allylic_oxid', 'fr_methoxy', 'fr_nh0', 'fr_coo2', 'fr_phenol', 'nacid', 'nbase', 'chi3v', 'fr_ar_nh', 'fr_nitrile', 'fr_imidazole', 'fr_urea', 'bcut2d_mrlow', 'chi1', 'smr_vsa6', 'fr_aryl_methyl', 'narombond', 'fr_alkyl_carbamate', 'fr_piperzine', 'exactmolwt', 'qed', 'chi0n', 'fr_sulfonamd', 'fr_thiazole', 'numvalenceelectrons', 'fr_phos_acid', 'peoe_vsa12', 'fr_nh1', 'fr_hdrzine', 'fr_c_o_nocoo', 'fr_lactone', 'estate_vsa6', 'bcut2d_logphi', 'vsa_estate7', 'peoe_vsa13', 'numsaturatedcarbocycles', 'fr_nitro', 'fr_phenol_noorthohbond', 'rotratio', 'fr_barbitur', 'fr_isothiocyan', 'balabanj', 'fr_arn', 'fr_imine', 'maxpartialcharge', 'fr_sulfide', 'slogp_vsa11', 'fr_hoccn', 'fr_n_o', 'peoe_vsa1', 'slogp_vsa6', 'heavyatommolwt', 'fractioncsp3', 'estate_vsa8', 'peoe_vsa10', 'numaliphaticrings', 'fr_thiophene', 'maxestateindex', 'smr_vsa10', 'labuteasa', 'smr_vsa2', 'fpdensitymorgan3', 'smr_vsa9', 'slogp_vsa10', 'numaromaticheterocycles', 'fr_nh2', 'fr_diazo', 'chi3n', 'fr_ar_coo', 'slogp_vsa5', 'fr_bicyclic', 'fr_amide', 'estate_vsa10', 'fr_guanido', 'chi1n', 'numsaturatedrings', 'fr_piperdine', 'fr_term_acetylene', 'estate_vsa4', 'slogp_vsa3', 'fr_coo', 'fr_ether', 'estate_vsa7', 'bcut2d_chglo', 'fr_oxazole', 'peoe_vsa6', 'hallkieralpha', 'peoe_vsa2', 'chi2v', 'nocount', 'vsa_estate5', 'fr_nhpyrrole', 'fr_al_coo', 'bertzct', 'estate_vsa11', 'minabspartialcharge', 'slogp_vsa8', 'fr_imide', 'kappa2', 'numaliphaticheterocycles', 'numsaturatedheterocycles', 'fr_hdrzone', 'smr_vsa4', 'fr_ar_n', 'nrot', 'smr_vsa8', 'slogp_vsa2', 'chi4v', 'fr_phos_ester', 'fr_para_hydroxylation', 'smr_vsa3', 'nhohcount', 'estate_vsa2', 'mollogp', 'tpsa', 'fr_azide', 'peoe_vsa4', 'numhdonors', 'fr_al_oh_notert', 'fr_c_o', 'chi0', 'fr_nitro_arom_nonortho', 'vsa_estate3', 'fr_benzene', 'fr_ketone', 'vsa_estate8', 'smr_vsa7', 'fr_ar_oh', 'fr_priamide', 'ringcount', 'estate_vsa3', 'numaromaticcarbocycles', 'bcut2d_mwhi', 'chi1v', 'heavyatomcount', 'vsa_estate4', 'chi0v', 'chiral_centers', 'r_cnt', 's_cnt', 'db_stereo', 'e_cnt', 'z_cnt', 'chiral_fp', 'db_fp'],
+    "compressed_features": [],
+    "train_all_data": True
 }
 
 
@@ -73,138 +69,99 @@ def match_features_case_insensitive(df: pd.DataFrame, model_features: list) -> p
     return df.rename(columns=rename_dict)
 
 
-def distance_weighted_calibrated_intervals(
-        df_pred: pd.DataFrame,
-        prox_df: pd.DataFrame,
-        calibration_strength: float = 0.7,
-        distance_decay: float = 3.0,
-) -> pd.DataFrame:
-    """
-    Calibrate intervals using distance-weighted neighbor quantiles.
-    Uses all 10 neighbors with distance-based weighting.
+def convert_categorical_types(df: pd.DataFrame, features: list, category_mappings={}) -> tuple:
     """
-    id_column = TEMPLATE_PARAMS["id_column"]
-    target_column = TEMPLATE_PARAMS["target"]
-
-    # Distance-weighted neighbor statistics
-    def weighted_quantile(values, weights, q):
-        """Calculate weighted quantile"""
-        if len(values) == 0:
-            return np.nan
-        sorted_indices = np.argsort(values)
-        sorted_values = values[sorted_indices]
-        sorted_weights = weights[sorted_indices]
-        cumsum = np.cumsum(sorted_weights)
-        cutoff = q * cumsum[-1]
-        return np.interp(cutoff, cumsum, sorted_values)
-
-    # Calculate distance weights (closer neighbors get more weight)
-    prox_df = prox_df.copy()
-    prox_df['weight'] = 1 / (1 + prox_df['distance'] ** distance_decay)
-
-    # Get weighted quantiles and statistics for each ID
-    neighbor_stats = []
-    for id_val, group in prox_df.groupby(id_column):
-        values = group[target_column].values
-        weights = group['weight'].values
-
-        # Normalize weights
-        weights = weights / weights.sum()
-
-        stats = {
-            id_column: id_val,
-            'local_q025': weighted_quantile(values, weights, 0.025),
-            'local_q25': weighted_quantile(values, weights, 0.25),
-            'local_q75': weighted_quantile(values, weights, 0.75),
-            'local_q975': weighted_quantile(values, weights, 0.975),
-            'local_median': weighted_quantile(values, weights, 0.5),
-            'local_std': np.sqrt(np.average((values - np.average(values, weights=weights)) ** 2, weights=weights)),
-            'avg_distance': group['distance'].mean(),
-            'min_distance': group['distance'].min(),
-            'max_distance': group['distance'].max(),
-        }
-        neighbor_stats.append(stats)
+    Converts appropriate columns to categorical type with consistent mappings.
 
-    neighbor_df = pd.DataFrame(neighbor_stats)
-    out = df_pred.merge(neighbor_df, on=id_column, how='left')
-
-    # Model disagreement score (normalized by prediction std)
-    model_disagreement = (out["prediction"] - out["prediction_uq"]).abs()
-    disagreement_score = (model_disagreement / out["prediction_std"]).clip(0, 2)
+    Args:
+        df (pd.DataFrame): The DataFrame to process.
+        features (list): List of feature names to consider for conversion.
+        category_mappings (dict, optional): Existing category mappings. If empty dict, we're in
+                                            training mode. If populated, we're in inference mode.
 
-    # Local confidence based on:
-    # 1. How close the neighbors are (closer = more confident)
-    # 2. How much local variance there is (less variance = more confident)
-    max_reasonable_distance = out['max_distance'].quantile(0.8)  # 80th percentile as reference
-    distance_confidence = (1 - (out['avg_distance'] / max_reasonable_distance)).clip(0.1, 1.0)
+    Returns:
+        tuple: (processed DataFrame, category mappings dictionary)
+    """
+    # Training mode
+    if category_mappings == {}:
+        for col in df.select_dtypes(include=["object", "string"]):
+            if col in features and df[col].nunique() < 20:
+                print(f"Training mode: Converting {col} to category")
+                df[col] = df[col].astype("category")
+                category_mappings[col] = df[col].cat.categories.tolist()  # Store category mappings
+
+    # Inference mode
+    else:
+        for col, categories in category_mappings.items():
+            if col in df.columns:
+                print(f"Inference mode: Applying categorical mapping for {col}")
+                df[col] = pd.Categorical(df[col], categories=categories)  # Apply consistent categorical mapping
 
-    variance_confidence = (out["prediction_std"] / out["local_std"]).clip(0.5, 2.0)
-    local_confidence = distance_confidence * variance_confidence.clip(0.5, 1.5)
+    return df, category_mappings
 
-    # Calibration weight: higher when models disagree and we have good local data
-    calibration_weight = (
-            calibration_strength *
-            local_confidence *  # Weight by local data quality
-            disagreement_score.clip(0.3, 1.0)  # More calibration when models disagree
-    )
 
-    # Consensus prediction (slight preference for NGBoost since it provides intervals)
-    consensus_pred = 0.65 * out["prediction_uq"] + 0.35 * out["prediction"]
+def decompress_features(
+        df: pd.DataFrame, features: List[str], compressed_features: List[str]
+) -> Tuple[pd.DataFrame, List[str]]:
+    """Prepare features for the model by decompressing bitstring features
 
-    # Re-center local intervals around consensus prediction
-    local_center_offset = consensus_pred - out["local_median"]
+    Args:
+        df (pd.DataFrame): The features DataFrame
+        features (List[str]): Full list of feature names
+        compressed_features (List[str]): List of feature names to decompress (bitstrings)
 
-    # Apply calibration to each quantile
-    quantile_pairs = [
-        ("q_025", "local_q025"),
-        ("q_25", "local_q25"),
-        ("q_75", "local_q75"),
-        ("q_975", "local_q975")
-    ]
+    Returns:
+        pd.DataFrame: DataFrame with the decompressed features
+        List[str]: Updated list of feature names after decompression
 
-    for model_q, local_q in quantile_pairs:
-        # Adjust local quantiles to be centered around consensus
-        adjusted_local_q = out[local_q] + local_center_offset
+    Raises:
+        ValueError: If any missing values are found in the specified features
+    """
 
-        # Blend model and local intervals
-        out[model_q] = (
-                (1 - calibration_weight) * out[model_q] +
-                calibration_weight * adjusted_local_q
+    # Check for any missing values in the required features
+    missing_counts = df[features].isna().sum()
+    if missing_counts.any():
+        missing_features = missing_counts[missing_counts > 0]
+        print(
+            f"WARNING: Found missing values in features: {missing_features.to_dict()}. "
+            "WARNING: You might want to remove/replace all NaN values before processing."
         )
 
-    # Ensure proper interval ordering and bounds using pandas
-    out["q_025"] = pd.concat([out["q_025"], consensus_pred], axis=1).min(axis=1)
-    out["q_975"] = pd.concat([out["q_975"], consensus_pred], axis=1).max(axis=1)
-    out["q_25"] = pd.concat([out["q_25"], out["q_75"]], axis=1).min(axis=1)
+    # Decompress the specified compressed features
+    decompressed_features = features.copy()
+    for feature in compressed_features:
+        if (feature not in df.columns) or (feature not in features):
+            print(f"Feature '{feature}' not in the features list, skipping decompression.")
+            continue
 
-    # Optional: Add some interval expansion when neighbors are very far
-    # (indicates we're in a sparse region of feature space)
-    sparse_region_mask = out['min_distance'] > out['min_distance'].quantile(0.9)
-    expansion_factor = 1 + 0.2 * sparse_region_mask  # 20% expansion in sparse regions
+        # Remove the feature from the list of features to avoid duplication
+        decompressed_features.remove(feature)
 
-    for q in ["q_025", "q_25", "q_75", "q_975"]:
-        interval_width = out[q] - consensus_pred
-        out[q] = consensus_pred + interval_width * expansion_factor
+        # Handle all compressed features as bitstrings
+        bit_matrix = np.array([list(bitstring) for bitstring in df[feature]], dtype=np.uint8)
+        prefix = feature[:3]
 
-    # Clean up temporary columns
-    cleanup_cols = [col for col in out.columns if col.startswith("local_")] + \
-                   ['avg_distance', 'min_distance', 'max_distance']
+        # Create all new columns at once - avoids fragmentation
+        new_col_names = [f"{prefix}_{i}" for i in range(bit_matrix.shape[1])]
+        new_df = pd.DataFrame(bit_matrix, columns=new_col_names, index=df.index)
 
-    return out.drop(columns=cleanup_cols)
+        # Add to features list
+        decompressed_features.extend(new_col_names)
+
+        # Drop original column and concatenate new ones
+        df = df.drop(columns=[feature])
+        df = pd.concat([df, new_df], axis=1)
+
+    return df, decompressed_features
 
 
-# TRAINING SECTION
-#
-# This section (__main__) is where SageMaker will execute the training job
-# and save the model artifacts to the model directory.
-#
 if __name__ == "__main__":
     # Template Parameters
-    id_column = TEMPLATE_PARAMS["id_column"]
-    features = TEMPLATE_PARAMS["features"]
     target = TEMPLATE_PARAMS["target"]
+    features = TEMPLATE_PARAMS["features"]
+    orig_features = features.copy()
+    compressed_features = TEMPLATE_PARAMS["compressed_features"]
     train_all_data = TEMPLATE_PARAMS["train_all_data"]
-    track_columns = TEMPLATE_PARAMS["track_columns"]  # Can be None
     validation_split = 0.2
 
     # Script arguments for input/output directories
@@ -216,102 +173,221 @@ if __name__ == "__main__":
     )
     args = parser.parse_args()
 
-    # Load training data from the specified directory
+    # Read the training data into DataFrames
     training_files = [
         os.path.join(args.train, file)
-        for file in os.listdir(args.train) if file.endswith(".csv")
+        for file in os.listdir(args.train)
+        if file.endswith(".csv")
     ]
     print(f"Training Files: {training_files}")
 
     # Combine files and read them all into a single pandas dataframe
-    df = pd.concat([pd.read_csv(file, engine="python") for file in training_files])
+    all_df = pd.concat([pd.read_csv(file, engine="python") for file in training_files])
+
+    # Check if the dataframe is empty
+    check_dataframe(all_df, "training_df")
+
+    # Features/Target output
+    print(f"Target: {target}")
+    print(f"Features: {str(features)}")
 
-    # Check if the DataFrame is empty
-    check_dataframe(df, "training_df")
+    # Convert any features that might be categorical to 'category' type
+    all_df, category_mappings = convert_categorical_types(all_df, features)
 
-    # Training data split logic
+    # If we have compressed features, decompress them
+    if compressed_features:
+        print(f"Decompressing features {compressed_features}...")
+        all_df, features = decompress_features(all_df, features, compressed_features)
+
+    # Do we want to train on all the data?
     if train_all_data:
-        # Use all data for both training and validation
-        print("Training on all data...")
-        df_train = df.copy()
-        df_val = df.copy()
-    elif "training" in df.columns:
-        # Split data based on a 'training' column if it exists
-        print("Splitting data based on 'training' column...")
-        df_train = df[df["training"]].copy()
-        df_val = df[~df["training"]].copy()
+        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"]]
+        df_val = all_df[~all_df["training"]]
     else:
-        # Perform a random split if no 'training' column is found
-        print("Splitting data randomly...")
-        df_train, df_val = train_test_split(df, test_size=validation_split, random_state=42)
-
-    # We're using XGBoost for point predictions and NGBoost for uncertainty quantification
-    xgb_model = XGBRegressor()
-    ngb_model = NGBRegressor()
+        # 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}")
 
     # Prepare features and targets for training
     X_train = df_train[features]
-    X_val = df_val[features]
+    X_validate = df_val[features]
     y_train = df_train[target]
-    y_val = df_val[target]
+    y_validate = df_val[target]
 
-    # Train both models using the training data
+    # Train XGBoost for point predictions
+    print("\nTraining XGBoost for point predictions...")
+    xgb_model = XGBRegressor(
+        n_estimators=1000,
+        max_depth=6,
+        learning_rate=0.01,
+        subsample=0.8,
+        colsample_bytree=0.8,
+        random_state=42,
+        verbosity=0
+    )
     xgb_model.fit(X_train, y_train)
-    ngb_model.fit(X_train, y_train, X_val=X_val, Y_val=y_val)
 
-    # Make Predictions on the Validation Set
-    print(f"Making Predictions on Validation Set...")
-    y_validate = df_val[target]
-    X_validate = df_val[features]
-    preds = xgb_model.predict(X_validate)
-
-    # Calculate various model performance metrics (regression)
-    rmse = root_mean_squared_error(y_validate, preds)
-    mae = mean_absolute_error(y_validate, preds)
-    r2 = r2_score(y_validate, preds)
-    print(f"RMSE: {rmse:.3f}")
-    print(f"MAE: {mae:.3f}")
-    print(f"R2: {r2:.3f}")
+    # Evaluate XGBoost performance
+    y_pred_xgb = xgb_model.predict(X_validate)
+    xgb_rmse = root_mean_squared_error(y_validate, y_pred_xgb)
+    xgb_mae = mean_absolute_error(y_validate, y_pred_xgb)
+    xgb_r2 = r2_score(y_validate, y_pred_xgb)
+
+    print(f"\nXGBoost Point Prediction Performance:")
+    print(f"RMSE: {xgb_rmse:.3f}")
+    print(f"MAE: {xgb_mae:.3f}")
+    print(f"R2: {xgb_r2:.3f}")
+
+    # Define confidence levels we want to model
+    confidence_levels = [0.50, 0.80, 0.90, 0.95]  # 50%, 80%, 90%, 95% confidence intervals
+
+    # Store MAPIE models for each confidence level
+    mapie_models = {}
+
+    # Train models for each confidence level
+    for confidence_level in confidence_levels:
+        alpha = 1 - confidence_level
+        lower_q = alpha / 2
+        upper_q = 1 - alpha / 2
+
+        print(f"\nTraining quantile models for {confidence_level * 100:.0f}% confidence interval...")
+        print(f"  Quantiles: {lower_q:.3f}, {upper_q:.3f}, 0.500")
+
+        # Train three models for this confidence level
+        quantile_estimators = []
+        for q in [lower_q, upper_q, 0.5]:
+            print(f"    Training model for quantile {q:.3f}...")
+            est = LGBMRegressor(
+                objective="quantile",
+                alpha=q,
+                n_estimators=1000,
+                max_depth=6,
+                learning_rate=0.01,
+                num_leaves=31,
+                min_child_samples=20,
+                subsample=0.8,
+                colsample_bytree=0.8,
+                random_state=42,
+                verbose=-1,
+                force_col_wise=True
+            )
+            est.fit(X_train, y_train)
+            quantile_estimators.append(est)
+
+        # Create MAPIE CQR model for this confidence level
+        print(f"  Setting up MAPIE CQR for {confidence_level * 100:.0f}% confidence...")
+        mapie_model = ConformalizedQuantileRegressor(
+            quantile_estimators,
+            confidence_level=confidence_level,
+            prefit=True
+        )
+
+        # Conformalize the model
+        print(f"  Conformalizing with validation data...")
+        mapie_model.conformalize(X_validate, y_validate)
+
+        # Store the model
+        mapie_models[f"mapie_{confidence_level:.2f}"] = mapie_model
+
+        # Validate coverage for this confidence level
+        y_pred, y_pis = mapie_model.predict_interval(X_validate)
+        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}%")
+
+    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)}")
 
+    # Analyze interval widths across confidence levels
+    print(f"\nInterval Width Analysis:")
+    for conf_level in confidence_levels:
+        model = mapie_models[f"mapie_{conf_level:.2f}"]
+        _, y_pis = model.predict_interval(X_validate)
+        widths = y_pis[:, 1, 0] - y_pis[:, 0, 0]
+        print(f"  {conf_level * 100:.0f}% CI: Mean width={np.mean(widths):.3f}, Std={np.std(widths):.3f}")
+
     # Save the trained XGBoost model
     xgb_model.save_model(os.path.join(args.model_dir, "xgb_model.json"))
 
-    # Save the trained NGBoost model
-    joblib.dump(ngb_model, os.path.join(args.model_dir, "ngb_model.joblib"))
+    # Save all MAPIE models
+    for model_name, model in mapie_models.items():
+        joblib.dump(model, os.path.join(args.model_dir, f"{model_name}.joblib"))
 
-    # Save the feature list to validate input during predictions
+    # Save the feature list
     with open(os.path.join(args.model_dir, "feature_columns.json"), "w") as fp:
         json.dump(features, fp)
 
-    # Now the Proximity model
-    model = Proximity(df_train, id_column, features, target, track_columns=track_columns)
+    # Save category mappings if any
+    if category_mappings:
+        with open(os.path.join(args.model_dir, "category_mappings.json"), "w") as fp:
+            json.dump(category_mappings, fp)
+
+    # Save model configuration
+    model_config = {
+        "model_type": "XGBoost_MAPIE_CQR_LightGBM",
+        "confidence_levels": confidence_levels,
+        "n_features": len(features),
+        "target": target,
+        "validation_metrics": {
+            "xgb_rmse": float(xgb_rmse),
+            "xgb_mae": float(xgb_mae),
+            "xgb_r2": float(xgb_r2),
+            "n_validation": len(df_val)
+        }
+    }
+    with open(os.path.join(args.model_dir, "model_config.json"), "w") as fp:
+        json.dump(model_config, fp, indent=2)
 
-    # Now serialize the model
-    model.serialize(args.model_dir)
+    print(f"\nModel training complete!")
+    print(f"Saved 1 XGBoost model and {len(mapie_models)} MAPIE models to {args.model_dir}")
 
 
 #
 # Inference Section
 #
 def model_fn(model_dir) -> dict:
-    """Load and return XGBoost and NGBoost regressors from model directory."""
+    """Load XGBoost and all MAPIE models from the specified directory."""
+
+    # Load model configuration to know which models to load
+    with open(os.path.join(model_dir, "model_config.json")) as fp:
+        config = json.load(fp)
 
     # Load XGBoost regressor
     xgb_path = os.path.join(model_dir, "xgb_model.json")
     xgb_model = XGBRegressor(enable_categorical=True)
     xgb_model.load_model(xgb_path)
 
-    # Load NGBoost regressor
-    ngb_model = joblib.load(os.path.join(model_dir, "ngb_model.joblib"))
+    # Load all MAPIE models
+    mapie_models = {}
+    for conf_level in config["confidence_levels"]:
+        model_name = f"mapie_{conf_level:.2f}"
+        mapie_models[model_name] = joblib.load(os.path.join(model_dir, f"{model_name}.joblib"))
 
-    # Deserialize the proximity model
-    prox_model = Proximity.deserialize(model_dir)
+    # Load category mappings if they exist
+    category_mappings = {}
+    category_path = os.path.join(model_dir, "category_mappings.json")
+    if os.path.exists(category_path):
+        with open(category_path) as fp:
+            category_mappings = json.load(fp)
 
     return {
-        "xgboost": xgb_model,
-        "ngboost": ngb_model,
-        "proximity": prox_model
+        "xgb_model": xgb_model,
+        "mapie_models": mapie_models,
+        "confidence_levels": config["confidence_levels"],
+        "category_mappings": category_mappings
     }
 
 
@@ -327,7 +403,7 @@ def input_fn(input_data, content_type):
     if "text/csv" in content_type:
         return pd.read_csv(StringIO(input_data))
     elif "application/json" in content_type:
-        return pd.DataFrame(json.loads(input_data))  # Assumes JSON array of records
+        return pd.DataFrame(json.loads(input_data))
     else:
         raise ValueError(f"{content_type} not supported!")
 
@@ -335,23 +411,26 @@ def input_fn(input_data, content_type):
 def output_fn(output_df, accept_type):
     """Supports both CSV and JSON output formats."""
     if "text/csv" in accept_type:
-        csv_output = output_df.fillna("N/A").to_csv(index=False)  # CSV with N/A for missing values
+        # Convert categorical columns to string to avoid fillna issues
+        for col in output_df.select_dtypes(include=['category']).columns:
+            output_df[col] = output_df[col].astype(str)
+        csv_output = output_df.fillna("N/A").to_csv(index=False)
         return csv_output, "text/csv"
     elif "application/json" in accept_type:
-        return output_df.to_json(orient="records"), "application/json"  # JSON array of records (NaNs -> null)
+        return output_df.to_json(orient="records"), "application/json"
     else:
         raise RuntimeError(f"{accept_type} accept type is not supported by this script.")
 
 
 def predict_fn(df, models) -> pd.DataFrame:
-    """Make Predictions with our XGB Quantile Regression Model
+    """Make predictions using XGBoost for point estimates and MAPIE for conformalized intervals
 
     Args:
         df (pd.DataFrame): The input DataFrame
-        models (dict): The dictionary of models to use for predictions
+        models (dict): Dictionary containing XGBoost and MAPIE models
 
     Returns:
-        pd.DataFrame: The DataFrame with the predictions added
+        pd.DataFrame: DataFrame with XGBoost predictions and conformalized intervals
     """
 
     # Grab our feature columns (from training)
@@ -362,32 +441,62 @@ def predict_fn(df, models) -> pd.DataFrame:
     # Match features in a case-insensitive manner
     matched_df = match_features_case_insensitive(df, model_features)
 
-    # Use XGBoost for point predictions
-    df["prediction"] = models["xgboost"].predict(matched_df[model_features])
-
-    # NGBoost predict returns distribution objects
-    y_dists = models["ngboost"].pred_dist(matched_df[model_features])
-
-    # Extract parameters from distribution
-    dist_params = y_dists.params
-
-    # Extract mean and std from distribution parameters
-    df["prediction_uq"] = dist_params['loc']  # mean
-    df["prediction_std"] = dist_params['scale']  # standard deviation
-
-    # Add 95% prediction intervals using ppf (percent point function)
-    df["q_025"] = y_dists.ppf(0.025)  # 2.5th percentile
-    df["q_975"] = y_dists.ppf(0.975)  # 97.5th percentile
-
-    # Add 50% prediction intervals
-    df["q_25"] = y_dists.ppf(0.25)   # 25th percentile
-    df["q_75"] = y_dists.ppf(0.75)   # 75th percentile
-
-    # Compute Nearest neighbors with Proximity model
-    prox_df = models["proximity"].neighbors(df)
+    # Apply categorical mappings if they exist
+    if models.get("category_mappings"):
+        matched_df, _ = convert_categorical_types(
+            matched_df,
+            model_features,
+            models["category_mappings"]
+        )
 
-    # Shrink prediction intervals based on KNN variance
-    df = distance_weighted_calibrated_intervals(df, prox_df)
+    # Get features for prediction
+    X = matched_df[model_features]
+
+    # Get XGBoost point predictions
+    df["prediction"] = models["xgb_model"].predict(X)
+
+    # Get predictions from each MAPIE model for conformalized intervals
+    for conf_level in models["confidence_levels"]:
+        model_name = f"mapie_{conf_level:.2f}"
+        model = models["mapie_models"][model_name]
+
+        # Get conformalized predictions
+        y_pred, y_pis = model.predict_interval(X)
+
+        # Map confidence levels to quantile names
+        if conf_level == 0.50:  # 50% CI
+            df["q_25"] = y_pis[:, 0, 0]
+            df["q_75"] = y_pis[:, 1, 0]
+        elif conf_level == 0.80:  # 80% CI
+            df["q_10"] = y_pis[:, 0, 0]
+            df["q_90"] = y_pis[:, 1, 0]
+        elif conf_level == 0.90:  # 90% CI
+            df["q_05"] = y_pis[:, 0, 0]
+            df["q_95"] = y_pis[:, 1, 0]
+        elif conf_level == 0.95:  # 95% CI
+            df["q_025"] = y_pis[:, 0, 0]
+            df["q_975"] = y_pis[:, 1, 0]
+
+    # Add median (q_50) from XGBoost prediction
+    df["q_50"] = df["prediction"]
+
+    # Calculate uncertainty metrics based on 95% interval
+    interval_width = df["q_975"] - df["q_025"]
+    df["prediction_std"] = interval_width / 3.92
+
+    # Reorder the quantile columns for easier reading
+    quantile_cols = ["q_025", "q_05", "q_10", "q_25", "q_75", "q_90", "q_95", "q_975"]
+    other_cols = [col for col in df.columns if col not in quantile_cols]
+    df = df[other_cols + quantile_cols]
+
+    # Uncertainty score
+    df["uncertainty_score"] = interval_width / (np.abs(df["prediction"]) + 1e-6)
+
+    # Confidence bands
+    df["confidence_band"] = pd.cut(
+        df["uncertainty_score"],
+        bins=[0, 0.5, 1.0, 2.0, np.inf],
+        labels=["high", "medium", "low", "very_low"]
+    )
 
-    # Return the modified DataFrame
     return df