workbench 0.8.202__py3-none-any.whl → 0.8.220__py3-none-any.whl

workbench/model_scripts/pytorch_model/generated_model_script.py

@@ -1,417 +1,265 @@
-# Imports for PyTorch Tabular Model
-import os
-import awswrangler as wr
-import numpy as np
-
-# PyTorch compatibility: pytorch-tabular saves complex objects, not just tensors
-# Use legacy loading behavior for compatibility (recommended by PyTorch docs for this scenario)
-os.environ["TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD"] = "1"
-from pytorch_tabular import TabularModel
-from pytorch_tabular.config import DataConfig, OptimizerConfig, TrainerConfig
-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
+# PyTorch Tabular Model Template for Workbench
+#
+# This template handles both classification and regression models with:
+# - K-fold cross-validation ensemble training (or single train/val split)
+# - Out-of-fold predictions for validation metrics
+# - Categorical feature embedding via TabularMLP
+# - Compressed feature decompression
+#
+# NOTE: Imports are structured to minimize serverless endpoint startup time.
+# Heavy imports (sklearn, awswrangler) are deferred to training time.
 
-# Classification Encoder
-from sklearn.preprocessing import LabelEncoder
-
-# Scikit Learn Imports
-from sklearn.model_selection import train_test_split, KFold, StratifiedKFold
-
-from io import StringIO
 import json
-import argparse
+import os
+
 import joblib
+import numpy as np
 import pandas as pd
+import torch
+
+from model_script_utils import (
+    convert_categorical_types,
+    decompress_features,
+    expand_proba_column,
+    input_fn,
+    match_features_case_insensitive,
+    output_fn,
+)
+from pytorch_utils import (
+    FeatureScaler,
+    load_model,
+    predict,
+    prepare_data,
+)
+from uq_harness import (
+    compute_confidence,
+    load_uq_models,
+    predict_intervals,
+)
 
-# Template Parameters
+# =============================================================================
+# Default Hyperparameters
+# =============================================================================
+DEFAULT_HYPERPARAMETERS = {
+    # Training parameters
+    "n_folds": 5,
+    "max_epochs": 200,
+    "early_stopping_patience": 30,
+    "batch_size": 128,
+    # Model architecture (larger capacity - ensemble provides regularization)
+    "layers": "512-256-128",
+    "learning_rate": 1e-3,
+    "dropout": 0.05,
+    "use_batch_norm": True,
+    # Loss function for regression (L1Loss=MAE, MSELoss=MSE, HuberLoss, SmoothL1Loss)
+    "loss": "L1Loss",
+    # Random seed
+    "seed": 42,
+}
+
+# Template parameters (filled in by Workbench)
 TEMPLATE_PARAMS = {
     "model_type": "uq_regressor",
     "target": "udm_asy_res_efflux_ratio",
-    "features": ['smr_vsa4', 'tpsa', 'nhohcount', 'peoe_vsa1', 'mollogp', 'numhdonors', 'tertiary_amine_count', 'smr_vsa3', 'nitrogen_span', 'vsa_estate2', 'hba_hbd_ratio', 'minpartialcharge', 'estate_vsa4', 'asphericity', 'charge_centroid_distance', 'peoe_vsa8', 'mi', 'estate_vsa8', 'vsa_estate6', 'vsa_estate3', 'molecular_volume_3d', 'kappa3', 'smr_vsa5', 'sv', 'xp_6dv', 'xc_4dv', 'si', 'molecular_axis_length', 'axp_5d', 'estate_vsa3', 'estate_vsa10', 'axp_7dv', 'slogp_vsa1', 'molecular_asymmetry', 'molmr', 'qed', 'xp_3d', 'axp_0dv', 'fpdensitymorgan1', 'minabsestateindex', 'numatomstereocenters', 'fpdensitymorgan2', 'slogp_vsa2', 'xch_5dv', 'num_s_centers', 'aromatic_interaction_score', 'axp_2dv', 'chi1v', 'hallkieralpha', 'vsa_estate8', 'peoe_vsa9', 'type_ii_pattern_count', 'slogp_vsa5', 'xc_3d', 'amphiphilic_moment', 'bcut2d_logphi', 'estate_vsa6', 'xc_3dv', 'chi0n', 'vsa_estate5', 'xpc_6d', 'vsa_estate7', 'axp_1d', 'axp_7d', 'xch_4dv', 'phi', 'maxestateindex', 'sps', 'bcut2d_mrlow', 'vsa_estate4', 'avgipc', 'bcut2d_mrhi', 'bcut2d_logplow', 'axp_1dv', 'kappa1', 'vsa_estate9', 'fr_imidazole', 'axp_6d', 'radius_of_gyration', 'chi2v', 'chi4n', 'xp_7d', 'smr_vsa6', 'axp_2d', 'num_r_centers', 'xch_7dv', 'estate_vsa2', 'axp_4d', 'maxpartialcharge', 'xpc_5dv', 'xp_5d', 'chi3n', 'chi2n', 'vsa_estate1', 'slogp_vsa11', 'bcut2d_mwlow', 'mm', 'c3sp3', 'numhacceptors', 'fr_nhpyrrole', 'labuteasa', 'fpdensitymorgan3', 'bcut2d_chghi', 'axp_3dv', 'c1sp3', 'kappa2', 'smr_vsa9', 'xp_6d', 'estate_vsa7', 'axp_6dv', 'sp', 'estate_vsa5', 'peoe_vsa2', 'smr_vsa1', 'mp', 'minestateindex', 'axp_3d', 'axp_4dv', 'chi0v', 'slogp_vsa3', 'heavyatommolwt', 'smr_vsa7', 'peoe_vsa6', 'mv', 'xp_4d', 'peoe_vsa7', 'mpe', 'chi4v', 'maxabspartialcharge', 'bcut2d_chglo', 'c1sp2', 'xp_0dv', 'smr_vsa10', 'estate_vsa1', 'fr_pyridine', 'bcut2d_mwhi', 'spe', 'balabanj', 'xch_7d', 'estate_vsa9', 'xp_3dv', 'fr_piperzine', 'xch_6dv', 'slogp_vsa8', 'peoe_vsa10', 'xp_4dv', 'c3sp2', 'fr_al_oh', 'xc_5d', 'fractioncsp3', 'fr_bicyclic', 'fr_piperdine', 'peoe_vsa12', 'peoe_vsa11', 'numheteroatoms', 'mse', 'xp_7dv', 'chi1', 'xpc_6dv', 'numsaturatedcarbocycles', 'chi1n', 'bertzct', 'xc_5dv', 'chi3v', 'intramolecular_hbond_potential', 'peoe_vsa4', 'xpc_5d', 'xp_2d', 'nbase', 'fr_priamide', 'slogp_vsa4', 'naromatom', 'vsa_estate10', 'fr_nitrile', 'molwt', 'peoe_vsa13', 'xch_4d', 'xp_5dv', 'numaromaticheterocycles', 'xpc_4dv', 'fr_hoccn', 'nocount', 'fr_nh1', 'mz', 'xc_6dv', 'hybratio', 'fr_imine', 'fr_morpholine', 'xpc_4d', 'xch_5d', 'numvalenceelectrons', 'numheterocycles', 'fr_aniline', 'fr_nh0', 'frac_defined_stereo', 'fr_benzene', 'xp_2dv', 'type_i_pattern_count', 'fr_ketone_topliss', 'fr_aryl_methyl', 'heavyatomcount', 'mare', 'axp_5dv', 'exactmolwt', 'xch_6d', 'xp_1d', 'xch_3d', 'axp_0d', 'amide_count', 'sse', 'slogp_vsa7', 'c2sp2', 'numrotatablebonds', 'chi0', 'xc_4d', 'slogp_vsa10', 'fr_al_oh_notert', 'numspiroatoms', 'numsaturatedrings', 'minabspartialcharge', 'fr_sulfone', 'slogp_vsa6', 'smr_vsa2', 'num_defined_stereocenters', 'numbridgeheadatoms', 'peoe_vsa3', 'numaliphaticheterocycles', 'fr_ndealkylation1', 'xc_6d'],
+    "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'],
     "id_column": "udm_mol_bat_id",
-    "compressed_features": [],
-    "model_metrics_s3_path": "s3://ideaya-sageworks-bucket/models/caco2-er-reg-pytorch-test/training",
-    "hyperparameters": {'n_folds': 5, 'training_config': {'max_epochs': 200, 'early_stopping_patience': 20}, 'model_config': {'layers': '256-128-64', 'dropout': 0.1, 'learning_rate': 0.001, 'activation': 'LeakyReLU'}},
+    "compressed_features": ['fingerprint'],
+    "model_metrics_s3_path": "s3://ideaya-sageworks-bucket/models/caco2-er-reg-temporal/training",
+    "hyperparameters": {'n_folds': 1},
 }
 
 
-def check_dataframe(df: pd.DataFrame, df_name: str) -> None:
-    """
-    Check if the provided dataframe is empty and raise an exception if it is.
-
-    Args:
-        df (pd.DataFrame): DataFrame to check
-        df_name (str): Name of the DataFrame
-    """
-    if df.empty:
-        msg = f"*** The training data {df_name} has 0 rows! ***STOPPING***"
-        print(msg)
-        raise ValueError(msg)
-
-
-def expand_proba_column(df: pd.DataFrame, class_labels: list[str]) -> pd.DataFrame:
-    """
-    Expands a column in a DataFrame containing a list of probabilities into separate columns.
-
-    Args:
-        df (pd.DataFrame): DataFrame containing a "pred_proba" column
-        class_labels (list[str]): List of class labels
-
-    Returns:
-        pd.DataFrame: DataFrame with the "pred_proba" expanded into separate columns
-    """
-    proba_column = "pred_proba"
-    if proba_column not in df.columns:
-        raise ValueError('DataFrame does not contain a "pred_proba" column')
-
-    # Construct new column names with '_proba' suffix
-    proba_splits = [f"{label}_proba" for label in class_labels]
-
-    # Expand the proba_column into separate columns for each probability
-    proba_df = pd.DataFrame(df[proba_column].tolist(), columns=proba_splits)
-
-    # Drop any proba columns and reset the index in prep for the concat
-    df = df.drop(columns=[proba_column] + proba_splits, errors="ignore")
-    df = df.reset_index(drop=True)
-
-    # Concatenate the new columns with the original DataFrame
-    df = pd.concat([df, proba_df], axis=1)
-    return df
-
-
-def match_features_case_insensitive(df: pd.DataFrame, model_features: list[str]) -> pd.DataFrame:
-    """
-    Matches and renames DataFrame columns to match model feature names (case-insensitive).
-    Prioritizes exact matches, then case-insensitive matches.
-
-    Raises ValueError if any model features cannot be matched.
-    """
-    df_columns_lower = {col.lower(): col for col in df.columns}
-    rename_dict = {}
-    missing = []
-    for feature in model_features:
-        if feature in df.columns:
-            continue  # Exact match
-        elif feature.lower() in df_columns_lower:
-            rename_dict[df_columns_lower[feature.lower()]] = feature
-        else:
-            missing.append(feature)
-
-    if missing:
-        raise ValueError(f"Features not found: {missing}")
-
-    # Rename the DataFrame columns to match the model features
-    return df.rename(columns=rename_dict)
-
-
-def convert_categorical_types(
-    df: pd.DataFrame, features: list[str], category_mappings: dict[str, list[str]] | None = None
-) -> tuple[pd.DataFrame, dict[str, list[str]]]:
-    """
-    Converts appropriate columns to categorical type with consistent mappings.
-
-    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 None or empty,
-                                            we're in training mode. If populated, we're in
-                                            inference mode.
-
-    Returns:
-        tuple: (processed DataFrame, category mappings dictionary)
-    """
-    if category_mappings is None:
-        category_mappings = {}
-
-    # Training mode
-    if not 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()
-
-    # 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)
-
-    return df, category_mappings
-
-
-def decompress_features(
-    df: pd.DataFrame, features: list[str], compressed_features: list[str]
-) -> tuple[pd.DataFrame, list[str]]:
-    """Prepare features for the model
-
-    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)
-
-    Returns:
-        pd.DataFrame: DataFrame with the decompressed features
-        list[str]: Updated list of feature names after decompression
-
-    Raises:
-        ValueError: If any missing values are found in the specified features
-    """
-    # 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."
-        )
-
-    # Make a copy to avoid mutating the original list
-    decompressed_features = features.copy()
-
-    for feature in compressed_features:
-        if (feature not in df.columns) or (feature not in decompressed_features):
-            print(f"Feature '{feature}' not in the features list, skipping decompression.")
-            continue
-
-        # Remove the feature from the list of features to avoid duplication
-        decompressed_features.remove(feature)
-
-        # Handle all compressed features as bitstrings
-        bit_matrix = np.array([list(bitstring) for bitstring in df[feature]], dtype=np.uint8)
-        prefix = feature[:3]
-
-        # 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)
-
-        # 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
-
-
+# =============================================================================
+# Model Loading (for SageMaker inference)
+# =============================================================================
 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(s)
-
-    Returns:
-        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"]
+    """Load PyTorch TabularMLP ensemble from the specified directory."""
+    # Load ensemble metadata
+    metadata_path = os.path.join(model_dir, "ensemble_metadata.joblib")
+    if os.path.exists(metadata_path):
+        metadata = joblib.load(metadata_path)
+        n_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)
+    # Determine device
+    device = "cuda" if torch.cuda.is_available() else "cpu"
 
-    # Save current working directory
-    original_cwd = os.getcwd()
+    # Load ensemble models
     ensemble_models = []
-
-    try:
-        # Change to /tmp because Pytorch Tabular needs write access (creates a .pt_tmp directory)
-        os.chdir("/tmp")
-
-        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 torch.load and working directory
-        torch.load = original_torch_load
-        os.chdir(original_cwd)
-
-    return {"ensemble_models": ensemble_models, "n_ensemble": n_ensemble}
-
-
-def input_fn(input_data, content_type: str) -> pd.DataFrame:
-    """Parse input data and return a DataFrame."""
-    if not input_data:
-        raise ValueError("Empty input data is not supported!")
-
-    # Decode bytes to string if necessary
-    if isinstance(input_data, bytes):
-        input_data = input_data.decode("utf-8")
-
-    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
-    else:
-        raise ValueError(f"{content_type} not supported!")
-
-
-def output_fn(output_df: pd.DataFrame, accept_type: str) -> tuple[str, str]:
-    """Supports both CSV and JSON output formats."""
-    if "text/csv" in accept_type:
-        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"
-    else:
-        raise RuntimeError(f"{accept_type} accept type is not supported by this script.")
+    for i in range(n_ensemble):
+        model_path = os.path.join(model_dir, f"model_{i}")
+        model = load_model(model_path, device=device)
+        ensemble_models.append(model)
+
+    print(f"Loaded {len(ensemble_models)} model(s)")
+
+    # Load feature scaler
+    scaler = FeatureScaler.load(os.path.join(model_dir, "scaler.joblib"))
+
+    # Load UQ models (regression only)
+    uq_models, uq_metadata = None, None
+    uq_path = os.path.join(model_dir, "uq_metadata.json")
+    if os.path.exists(uq_path):
+        uq_models, uq_metadata = load_uq_models(model_dir)
+
+    return {
+        "ensemble_models": ensemble_models,
+        "n_ensemble": n_ensemble,
+        "scaler": scaler,
+        "uq_models": uq_models,
+        "uq_metadata": uq_metadata,
+    }
 
 
+# =============================================================================
+# Inference (for SageMaker inference)
+# =============================================================================
 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_dict: Dictionary containing ensemble models and metadata
-
-    Returns:
-        pd.DataFrame: The DataFrame with predictions (and prediction_std for ensembles)
-    """
+    """Make predictions with PyTorch TabularMLP ensemble."""
     model_type = TEMPLATE_PARAMS["model_type"]
     compressed_features = TEMPLATE_PARAMS["compressed_features"]
+    model_dir = os.environ.get("SM_MODEL_DIR", "/opt/ml/model")
 
-    # Extract ensemble models
+    # Load artifacts
     ensemble_models = model_dict["ensemble_models"]
-    n_ensemble = model_dict["n_ensemble"]
+    scaler = model_dict["scaler"]
+    uq_models = model_dict.get("uq_models")
+    uq_metadata = model_dict.get("uq_metadata")
 
-    # 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:
-        features = json.load(fp)
-    print(f"Model Features: {features}")
+    with open(os.path.join(model_dir, "feature_columns.json")) as f:
+        features = json.load(f)
+    with open(os.path.join(model_dir, "category_mappings.json")) as f:
+        category_mappings = json.load(f)
+    with open(os.path.join(model_dir, "feature_metadata.json")) as f:
+        feature_metadata = json.load(f)
 
-    # Load the category mappings (from training)
-    with open(os.path.join(model_dir, "category_mappings.json")) as fp:
-        category_mappings = json.load(fp)
+    continuous_cols = feature_metadata["continuous_cols"]
+    categorical_cols = feature_metadata["categorical_cols"]
 
-    # Load our Label Encoder if we have one
     label_encoder = None
-    label_encoder_path = os.path.join(model_dir, "label_encoder.joblib")
-    if os.path.exists(label_encoder_path):
-        label_encoder = joblib.load(label_encoder_path)
+    encoder_path = os.path.join(model_dir, "label_encoder.joblib")
+    if os.path.exists(encoder_path):
+        label_encoder = joblib.load(encoder_path)
 
-    # Match features in a case-insensitive manner
-    matched_df = match_features_case_insensitive(df, features)
+    print(f"Model Features: {features}")
 
-    # Detect categorical types in the incoming DataFrame
+    # Prepare features
+    matched_df = match_features_case_insensitive(df, features)
     matched_df, _ = convert_categorical_types(matched_df, features, category_mappings)
 
-    # If we have compressed features, decompress them
     if compressed_features:
         print("Decompressing features for prediction...")
         matched_df, features = decompress_features(matched_df, features, compressed_features)
 
-    # Track rows with missing features
+    # Track missing features
     missing_mask = matched_df[features].isna().any(axis=1)
     if missing_mask.any():
-        print(f"Warning: {missing_mask.sum()} rows have missing features, will return NaN predictions")
+        print(f"Warning: {missing_mask.sum()} rows have missing features")
 
-    # Initialize prediction columns
+    # Initialize output 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]
+    complete_df = matched_df[~missing_mask].copy()
     if len(complete_df) == 0:
         print("Warning: No complete rows to predict on")
         return df
 
-    # pytorch-tabular returns predictions using f"{target}_prediction" column
-    target = TEMPLATE_PARAMS["target"]
-    prediction_column = f"{target}_prediction"
-
-    # Collect predictions from all ensemble members
-    all_ensemble_preds = []
-    all_ensemble_probs = []
-
-    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()}")
-
-        all_ensemble_preds.append(ens_preds)
+    # Prepare data for inference (with standardization)
+    x_cont, x_cat, _, _, _ = prepare_data(
+        complete_df, continuous_cols, categorical_cols, category_mappings=category_mappings, scaler=scaler
+    )
 
-        # 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)
+    # Collect ensemble predictions
+    all_preds = []
+    for model in ensemble_models:
+        preds = predict(model, x_cont, x_cat)
+        all_preds.append(preds)
 
-    # Stack and compute mean/std (std is 0 for single model)
-    ensemble_preds = np.stack(all_ensemble_preds, axis=0)  # (n_ensemble, n_samples)
+    # Aggregate predictions
+    ensemble_preds = np.stack(all_preds, axis=0)
     preds = np.mean(ensemble_preds, axis=0)
-    preds_std = np.std(ensemble_preds, axis=0)  # Will be 0s for n_ensemble=1
+    preds_std = np.std(ensemble_preds, axis=0)
 
-    print(f"Inference: Ensemble predictions shape: {preds.shape}, n_ensemble: {n_ensemble}")
+    print(f"Inference complete: {len(preds)} predictions, {len(ensemble_models)} ensemble members")
 
-    # Handle classification vs regression
     if label_encoder is not None:
-        # 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 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))
+        # Classification: average probabilities, then argmax
+        avg_probs = preds  # Already softmax output
+        class_preds = np.argmax(avg_probs, axis=1)
+        predictions = label_encoder.inverse_transform(class_preds)
+
+        all_proba = pd.Series([None] * len(df), index=df.index, dtype=object)
+        all_proba.loc[~missing_mask] = [p.tolist() for p in avg_probs]
+        df["pred_proba"] = all_proba
+        df = expand_proba_column(df, label_encoder.classes_)
     else:
-        # Regression (includes uq_regressor)
-        predictions = preds
-        df.loc[~missing_mask, "prediction_std"] = preds_std
+        # Regression
+        predictions = preds.flatten()
+        df.loc[~missing_mask, "prediction_std"] = preds_std.flatten()
+
+        # Add UQ intervals if available
+        if uq_models and uq_metadata:
+            X_complete = complete_df[features]
+            df_complete = df.loc[~missing_mask].copy()
+            df_complete["prediction"] = predictions  # Set prediction before compute_confidence
+            df_complete = predict_intervals(df_complete, X_complete, uq_models, uq_metadata)
+            df_complete = compute_confidence(df_complete, uq_metadata["median_interval_width"], "q_10", "q_90")
+            # Copy UQ columns back to main dataframe
+            for col in df_complete.columns:
+                if col.startswith("q_") or col == "confidence":
+                    df.loc[~missing_mask, col] = df_complete[col].values
 
-    # Set predictions only for complete rows
     df.loc[~missing_mask, "prediction"] = predictions
-
     return df
 
 
+# =============================================================================
+# Training
+# =============================================================================
 if __name__ == "__main__":
-    """The main function is for training the PyTorch Tabular model"""
+    # -------------------------------------------------------------------------
+    # Training-only imports (deferred to reduce serverless startup time)
+    # -------------------------------------------------------------------------
+    import argparse
+
+    import awswrangler as wr
+    from sklearn.model_selection import KFold, StratifiedKFold, train_test_split
+    from sklearn.preprocessing import LabelEncoder
+
+    # Enable Tensor Core optimization for GPUs that support it
+    torch.set_float32_matmul_precision("medium")
+
+    from model_script_utils import (
+        check_dataframe,
+        compute_classification_metrics,
+        compute_regression_metrics,
+        print_classification_metrics,
+        print_confusion_matrix,
+        print_regression_metrics,
+    )
+    from pytorch_utils import (
+        create_model,
+        save_model,
+        train_model,
+    )
+    from uq_harness import (
+        save_uq_models,
+        train_uq_models,
+    )
+
+    # -------------------------------------------------------------------------
+    # Setup: Parse arguments and load data
+    # -------------------------------------------------------------------------
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--model-dir", type=str, default=os.environ.get("SM_MODEL_DIR", "/opt/ml/model"))
+    parser.add_argument("--train", type=str, default=os.environ.get("SM_CHANNEL_TRAIN", "/opt/ml/input/data/train"))
+    parser.add_argument("--output-data-dir", type=str, default=os.environ.get("SM_OUTPUT_DATA_DIR", "/opt/ml/output/data"))
+    args = parser.parse_args()
 
-    # Harness Template Parameters
+    # Extract template parameters
     target = TEMPLATE_PARAMS["target"]
     features = TEMPLATE_PARAMS["features"]
     orig_features = features.copy()
@@ -419,341 +267,254 @@ if __name__ == "__main__":
     compressed_features = TEMPLATE_PARAMS["compressed_features"]
     model_type = TEMPLATE_PARAMS["model_type"]
     model_metrics_s3_path = TEMPLATE_PARAMS["model_metrics_s3_path"]
-    hyperparameters = TEMPLATE_PARAMS["hyperparameters"]
+    hyperparameters = {**DEFAULT_HYPERPARAMETERS, **(TEMPLATE_PARAMS["hyperparameters"] or {})}
 
-    # Script arguments for input/output directories
-    parser = argparse.ArgumentParser()
-    parser.add_argument("--model-dir", type=str, default=os.environ.get("SM_MODEL_DIR", "/opt/ml/model"))
-    parser.add_argument("--train", type=str, default=os.environ.get("SM_CHANNEL_TRAIN", "/opt/ml/input/data/train"))
-    parser.add_argument(
-        "--output-data-dir", type=str, default=os.environ.get("SM_OUTPUT_DATA_DIR", "/opt/ml/output/data")
-    )
-    args = parser.parse_args()
-
-    # 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")]
+    # Load training data
+    training_files = [os.path.join(args.train, f) for f in os.listdir(args.train) if f.endswith(".csv")]
     print(f"Training Files: {training_files}")
-
-    # Combine files and read them all into a single pandas dataframe
-    all_df = pd.concat([pd.read_csv(file, engine="python") for file in training_files])
-
-    # Print out some info about the dataframe
-    print(f"All Data Shape: {all_df.shape}")
-    print(f"Feature dtypes:\n{all_df[features].dtypes.value_counts()}")
-    print(f"Int64 columns: {all_df[features].select_dtypes(include=['int64']).columns.tolist()}")
-
-    # Check if the dataframe is empty
+    all_df = pd.concat([pd.read_csv(f, engine="python") for f in training_files])
     check_dataframe(all_df, "training_df")
 
-    # Drop any rows with missing feature values
-    initial_row_count = all_df.shape[0]
+    # Drop rows with missing features
+    initial_count = len(all_df)
     all_df = all_df.dropna(subset=features)
-    dropped_rows = initial_row_count - all_df.shape[0]
-    if dropped_rows > 0:
-        print(f"Dropped {dropped_rows} rows due to missing feature values.")
+    if len(all_df) < initial_count:
+        print(f"Dropped {initial_count - len(all_df)} rows with missing features")
 
-    # Features/Target output
     print(f"Target: {target}")
-    print(f"Features: {str(features)}")
+    print(f"Features: {features}")
+    print(f"Hyperparameters: {hyperparameters}")
 
-    # Convert any features that might be categorical to 'category' type
+    # -------------------------------------------------------------------------
+    # Preprocessing
+    # -------------------------------------------------------------------------
     all_df, category_mappings = convert_categorical_types(all_df, features)
 
-    # Print out some info about the dataframe
-    print(f"All Data Shape: {all_df.shape}")
-    print(f"Feature dtypes:\n{all_df[features].dtypes.value_counts()}")
-    print(f"Int64 columns: {all_df[features].select_dtypes(include=['int64']).columns.tolist()}")
-
-    # If we have compressed features, decompress them
     if compressed_features:
-        print(f"Decompressing features {compressed_features}...")
+        print(f"Decompressing features: {compressed_features}")
         all_df, features = decompress_features(all_df, features, compressed_features)
 
-    # Determine categorical and continuous columns
-    categorical_cols = [col for col in features if all_df[col].dtype.name == "category"]
-    continuous_cols = [col for col in features if col not in categorical_cols]
-    print(f"Categorical columns: {categorical_cols}")
-    print(f"Continuous columns: {continuous_cols}")
-
-    # Cast continuous columns to float
+    # Determine categorical vs continuous columns
+    categorical_cols = [c for c in features if all_df[c].dtype.name == "category"]
+    continuous_cols = [c for c in features if c not in categorical_cols]
     all_df[continuous_cols] = all_df[continuous_cols].astype("float64")
+    print(f"Categorical: {categorical_cols}")
+    print(f"Continuous: {len(continuous_cols)} columns")
 
-    # Choose the 'task' based on model type and set up the label encoder if needed
+    # -------------------------------------------------------------------------
+    # Classification setup
+    # -------------------------------------------------------------------------
+    label_encoder = None
+    n_outputs = 1
     if model_type == "classifier":
-        task = "classification"
-        # Encode the target column on full dataset for consistent encoding
         label_encoder = LabelEncoder()
         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
+        n_outputs = len(label_encoder.classes_)
+        print(f"Class labels: {label_encoder.classes_.tolist()}")
 
-    # 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)
+    # -------------------------------------------------------------------------
+    # Cross-validation setup
+    # -------------------------------------------------------------------------
+    n_folds = hyperparameters["n_folds"]
+    task = "classification" if model_type == "classifier" else "regression"
+    hidden_layers = [int(x) for x in hyperparameters["layers"].split("-")]
 
-    # =========================================================================
-    # 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'}...")
+    # Get categorical cardinalities
+    categorical_cardinalities = [len(category_mappings.get(col, {})) for col in categorical_cols]
 
-    # 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")
+            print("Using 'training' column for train/val split")
             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)
+            print("WARNING: No 'training' column found, using random 80/20 split")
+            train_idx, val_idx = train_test_split(np.arange(len(all_df)), 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]
+            folds = list(kfold.split(all_df, 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
+            folds = list(kfold.split(all_df))
 
-    ensemble_models = []
+    print(f"Training {'single model' if n_folds == 1 else f'{n_folds}-fold ensemble'}...")
 
-    # Set up PyTorch Tabular data configuration (shared across folds)
-    data_config = DataConfig(
-        target=[target],
-        continuous_cols=continuous_cols,
-        categorical_cols=categorical_cols,
-    )
+    # Fit scaler on all training data (used across all folds)
+    scaler = FeatureScaler()
+    scaler.fit(all_df, continuous_cols)
+    print(f"Fitted scaler on {len(continuous_cols)} continuous features")
 
-    # Model config defaults
-    model_defaults = {
-        "layers": "256-128-64",
-        "activation": "LeakyReLU",
-        "learning_rate": 1e-3,
-        "dropout": 0.1,
-        "use_batch_norm": True,
-        "initialization": "kaiming",
-    }
-    # Override defaults with model_config if present
-    model_overrides = {k: v for k, v in hyperparameters.get("model_config", {}).items() if k in model_defaults}
-    for key, value in model_overrides.items():
-        print(f"MODEL CONFIG Override: {key}: {model_defaults[key]} → {value}")
-    model_params = {**model_defaults, **model_overrides}
+    # Determine device
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    print(f"Using device: {device}")
 
-    model_config = CategoryEmbeddingModelConfig(task=task, **model_params)
-    optimizer_config = OptimizerConfig()
+    # -------------------------------------------------------------------------
+    # Training loop
+    # -------------------------------------------------------------------------
+    oof_predictions = np.full((len(all_df), n_outputs), np.nan, dtype=np.float64)
 
+    ensemble_models = []
     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"Fold {fold_idx + 1}/{len(folds)} - Train: {len(train_idx)}, Val: {len(val_idx)}")
         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,
+        # Prepare data (using pre-fitted scaler)
+        train_x_cont, train_x_cat, train_y, _, _ = prepare_data(
+            df_train, continuous_cols, categorical_cols, target, category_mappings, scaler=scaler
+        )
+        val_x_cont, val_x_cat, val_y, _, _ = prepare_data(
+            df_val, continuous_cols, categorical_cols, target, category_mappings, scaler=scaler
         )
-        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
+        # Create model
+        torch.manual_seed(hyperparameters["seed"] + fold_idx)
+        model = create_model(
+            n_continuous=len(continuous_cols),
+            categorical_cardinalities=categorical_cardinalities,
+            hidden_layers=hidden_layers,
+            n_outputs=n_outputs,
+            task=task,
+            dropout=hyperparameters["dropout"],
+            use_batch_norm=hyperparameters["use_batch_norm"],
+        )
 
-        # 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()
+        # Train
+        model, history = train_model(
+            model,
+            train_x_cont, train_x_cat, train_y,
+            val_x_cont, val_x_cat, val_y,
+            task=task,
+            max_epochs=hyperparameters["max_epochs"],
+            patience=hyperparameters["early_stopping_patience"],
+            batch_size=hyperparameters["batch_size"],
+            learning_rate=hyperparameters["learning_rate"],
+            loss=hyperparameters.get("loss", "L1Loss"),
+            device=device,
+        )
+        ensemble_models.append(model)
 
-        print(f"Fold {fold_idx + 1} complete!")
+        # Out-of-fold predictions
+        fold_preds = predict(model, val_x_cont, val_x_cat)
+        oof_predictions[val_idx] = fold_preds
 
     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
+    # -------------------------------------------------------------------------
+    # Prepare validation results
+    # -------------------------------------------------------------------------
     if n_folds == 1:
-        val_mask = ~np.isnan(oof_predictions)
-        preds = oof_predictions[val_mask]
+        val_mask = ~np.isnan(oof_predictions[:, 0])
         df_val = all_df[val_mask].copy()
-        if oof_proba is not None:
-            oof_proba = oof_proba[val_mask]
+        predictions = oof_predictions[val_mask]
     else:
-        preds = oof_predictions
         df_val = all_df.copy()
+        predictions = oof_predictions
+
+    # Decode labels for classification
+    if model_type == "classifier":
+        class_preds = np.argmax(predictions, axis=1)
+        df_val[target] = label_encoder.inverse_transform(df_val[target].astype(int))
+        df_val["prediction"] = label_encoder.inverse_transform(class_preds)
+        df_val["pred_proba"] = [p.tolist() for p in predictions]
+        df_val = expand_proba_column(df_val, label_encoder.classes_)
+    else:
+        df_val["prediction"] = predictions.flatten()
 
-    # 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}")
+    # -------------------------------------------------------------------------
+    # Compute and print metrics
+    # -------------------------------------------------------------------------
+    y_true = df_val[target].values
+    y_pred = df_val["prediction"].values
 
     if model_type == "classifier":
-        # Get probabilities for classification
-        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_)
-
-        # Decode the target and prediction labels
-        y_validate = label_encoder.inverse_transform(df_val[target])
-        preds_decoded = label_encoder.inverse_transform(preds.astype(int))
+        score_df = compute_classification_metrics(y_true, y_pred, label_encoder.classes_, target)
+        print_classification_metrics(score_df, target, label_encoder.classes_)
+        print_confusion_matrix(y_true, y_pred, label_encoder.classes_)
     else:
-        y_validate = df_val[target].values
-        preds_decoded = preds
+        metrics = compute_regression_metrics(y_true, y_pred)
+        print_regression_metrics(metrics)
+
+        # Compute ensemble prediction_std
+        if n_folds > 1:
+            # Re-run inference with all models to get std
+            x_cont, x_cat, _, _, _ = prepare_data(
+                df_val, continuous_cols, categorical_cols, category_mappings=category_mappings, scaler=scaler
+            )
+            all_preds = [predict(m, x_cont, x_cat).flatten() for m in ensemble_models]
+            df_val["prediction_std"] = np.std(np.stack(all_preds), axis=0)
+            print(f"Ensemble std - mean: {df_val['prediction_std'].mean():.4f}, max: {df_val['prediction_std'].max():.4f}")
+        else:
+            df_val["prediction_std"] = 0.0
 
-    # Save predictions to S3
-    df_val = df_val.copy()
-    df_val["prediction"] = preds_decoded
+        # Train UQ models for uncertainty quantification
+        print("\n" + "=" * 50)
+        print("Training UQ Models")
+        print("=" * 50)
+        uq_models, uq_metadata = train_uq_models(
+            all_df[features], all_df[target], df_val[features], y_true
+        )
+        df_val = predict_intervals(df_val, df_val[features], uq_models, uq_metadata)
+        df_val = compute_confidence(df_val, uq_metadata["median_interval_width"])
 
-    # Build output columns - include id_column if it exists
+    # -------------------------------------------------------------------------
+    # Save validation predictions to S3
+    # -------------------------------------------------------------------------
     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
+    if model_type != "classifier":
         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 += [c for c in df_val.columns if c.startswith("q_") or c == "confidence"]
+
+    output_columns += [c for c in df_val.columns if c.endswith("_proba")]
+
+    wr.s3.to_csv(df_val[output_columns], f"{model_metrics_s3_path}/validation_predictions.csv", index=False)
+
+    # -------------------------------------------------------------------------
+    # Save model artifacts
+    # -------------------------------------------------------------------------
+    model_config = {
+        "n_continuous": len(continuous_cols),
+        "categorical_cardinalities": categorical_cardinalities,
+        "hidden_layers": hidden_layers,
+        "n_outputs": n_outputs,
+        "task": task,
+        "dropout": hyperparameters["dropout"],
+        "use_batch_norm": hyperparameters["use_batch_norm"],
+    }
 
-    output_columns += [col for col in df_val.columns if col.endswith("_proba")]
-    wr.s3.to_csv(
-        df_val[output_columns],
-        path=f"{model_metrics_s3_path}/validation_predictions.csv",
-        index=False,
-    )
+    for idx, m in enumerate(ensemble_models):
+        save_model(m, os.path.join(args.model_dir, f"model_{idx}"), model_config)
+    print(f"Saved {len(ensemble_models)} model(s)")
 
-    # Report Performance Metrics
-    if model_type == "classifier":
-        # Get the label names and their integer mapping
-        label_names = label_encoder.classes_
-
-        # Calculate various model performance metrics
-        scores = precision_recall_fscore_support(y_validate, preds_decoded, average=None, labels=label_names)
-
-        # Put the scores into a dataframe
-        score_df = pd.DataFrame(
-            {
-                target: label_names,
-                "precision": scores[0],
-                "recall": scores[1],
-                "f1": scores[2],
-                "support": scores[3],
-            }
-        )
+    joblib.dump({"n_ensemble": len(ensemble_models), "n_folds": n_folds}, os.path.join(args.model_dir, "ensemble_metadata.joblib"))
 
-        # Output metrics per class
-        metrics = ["precision", "recall", "f1", "support"]
-        for t in label_names:
-            for m in metrics:
-                value = score_df.loc[score_df[target] == t, m].iloc[0]
-                print(f"Metrics:{t}:{m} {value}")
+    with open(os.path.join(args.model_dir, "feature_columns.json"), "w") as f:
+        json.dump(orig_features, f)
 
-        # Compute and output the confusion matrix
-        conf_mtx = confusion_matrix(y_validate, preds_decoded, labels=label_names)
-        for i, row_name in enumerate(label_names):
-            for j, col_name in enumerate(label_names):
-                value = conf_mtx[i, j]
-                print(f"ConfusionMatrix:{row_name}:{col_name} {value}")
+    with open(os.path.join(args.model_dir, "category_mappings.json"), "w") as f:
+        json.dump(category_mappings, f)
 
-    else:
-        # 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)
-        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})")
+    with open(os.path.join(args.model_dir, "feature_metadata.json"), "w") as f:
+        json.dump({"continuous_cols": continuous_cols, "categorical_cols": categorical_cols}, f)
+
+    with open(os.path.join(args.model_dir, "hyperparameters.json"), "w") as f:
+        json.dump(hyperparameters, f, indent=2)
+
+    scaler.save(os.path.join(args.model_dir, "scaler.joblib"))
 
     if label_encoder:
         joblib.dump(label_encoder, os.path.join(args.model_dir, "label_encoder.joblib"))
 
-    # Save the features (this will validate input during predictions)
-    with open(os.path.join(args.model_dir, "feature_columns.json"), "w") as fp:
-        json.dump(orig_features, fp)
+    if model_type != "classifier":
+        save_uq_models(uq_models, uq_metadata, args.model_dir)
 
-    # Save the category mappings
-    with open(os.path.join(args.model_dir, "category_mappings.json"), "w") as fp:
-        json.dump(category_mappings, fp)
+    print(f"\nModel training complete! Artifacts saved to {args.model_dir}")