workbench 0.8.160__py3-none-any.whl → 0.8.202__py3-none-any.whl

workbench/model_scripts/pytorch_model/generated_model_script.py

@@ -8,43 +8,43 @@ import numpy as np
 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, TabNetModelConfig
+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
 import argparse
 import joblib
-import os
 import pandas as pd
-from typing import List, Tuple
 
 # Template Parameters
 TEMPLATE_PARAMS = {
-    "model_type": "classifier",
-    "target_column": "solubility_class",
-    "features": ['molwt', 'mollogp', 'molmr', 'heavyatomcount', 'numhacceptors', 'numhdonors', 'numheteroatoms', 'numrotatablebonds', 'numvalenceelectrons', 'numaromaticrings', 'numsaturatedrings', 'numaliphaticrings', 'ringcount', 'tpsa', 'labuteasa', 'balabanj', 'bertzct'],
+    "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'],
+    "id_column": "udm_mol_bat_id",
     "compressed_features": [],
-    "model_metrics_s3_path": "s3://sandbox-sageworks-artifacts/models/aqsol-pytorch-class/training",
-    "train_all_data": False
+    "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'}},
 }
 
 
-# Function to check if dataframe is empty
 def check_dataframe(df: pd.DataFrame, df_name: str) -> None:
     """
     Check if the provided dataframe is empty and raise an exception if it is.
@@ -59,19 +59,17 @@ def check_dataframe(df: pd.DataFrame, df_name: str) -> None:
         raise ValueError(msg)
 
 
-def expand_proba_column(df: pd.DataFrame, class_labels: List[str]) -> pd.DataFrame:
+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
+        class_labels (list[str]): List of class labels
 
     Returns:
         pd.DataFrame: DataFrame with the "pred_proba" expanded into separate columns
     """
-
-    # Sanity check
     proba_column = "pred_proba"
     if proba_column not in df.columns:
         raise ValueError('DataFrame does not contain a "pred_proba" column')
@@ -88,11 +86,10 @@ def expand_proba_column(df: pd.DataFrame, class_labels: List[str]) -> pd.DataFra
 
     # Concatenate the new columns with the original DataFrame
     df = pd.concat([df, proba_df], axis=1)
-    print(df)
     return df
 
 
-def match_features_case_insensitive(df: pd.DataFrame, model_features: list) -> pd.DataFrame:
+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.
@@ -102,7 +99,6 @@ def match_features_case_insensitive(df: pd.DataFrame, model_features: list) -> p
     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
@@ -114,58 +110,64 @@ def match_features_case_insensitive(df: pd.DataFrame, model_features: list) -> p
     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, category_mappings={}) -> tuple:
+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 empty dict, we're in
-                                            training mode. If populated, we're in inference mode.
+        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 category_mappings == {}:
+    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()  # Store category mappings
+                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)  # Apply consistent categorical mapping
+                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]]:
+    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)
+        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
+        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():
@@ -175,10 +177,11 @@ def decompress_features(
             "WARNING: You might want to remove/replace all NaN values before processing."
         )
 
-    # Decompress the specified compressed features
-    decompressed_features = features
+    # 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 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
 
@@ -203,35 +206,60 @@ def decompress_features(
     return df, decompressed_features
 
 
-def model_fn(model_dir):
-    """Deserialize and return fitted PyTorch Tabular model"""
-    #
-    os.environ['TEMP'] = '/tmp'
-    model_path = os.path.join(model_dir, "tabular_model")
-    model = TabularModel.load_model(model_path)
-    return model
+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)
 
-def model_fn(model_dir):
+    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"]
+    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')
+        os.chdir("/tmp")
 
-        # Load the model
-        model_path = os.path.join(model_dir, "tabular_model")
-        model = TabularModel.load_model(model_path)
+        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)
 
-    # Restore the original working directory
     finally:
+        # 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):
+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!")
@@ -248,29 +276,34 @@ def input_fn(input_data, content_type):
         raise ValueError(f"{content_type} not supported!")
 
 
-def output_fn(output_df, accept_type):
+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)  # CSV with N/A for missing values
+        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, model) -> 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:
@@ -283,12 +316,11 @@ def predict_fn(df, model) -> pd.DataFrame:
 
     # Load our Label Encoder if we have one
     label_encoder = None
-    if os.path.exists(os.path.join(model_dir, "label_encoder.joblib")):
-        label_encoder = joblib.load(os.path.join(model_dir, "label_encoder.joblib"))
+    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)
 
-    # We're going match features in a case-insensitive manner, accounting for all the permutations
-    # - Model has a feature list that's any case ("Id", "taCos", "cOunT", "likes_tacos")
-    # - Incoming data has columns that are mixed case ("ID", "Tacos", "Count", "Likes_Tacos")
+    # Match features in a case-insensitive manner
     matched_df = match_features_case_insensitive(df, features)
 
     # Detect categorical types in the incoming DataFrame
@@ -299,36 +331,80 @@ def predict_fn(df, model) -> pd.DataFrame:
         print("Decompressing features for prediction...")
         matched_df, features = decompress_features(matched_df, features, compressed_features)
 
-    # Make predictions using the TabularModel
-    result = model.predict(matched_df[features])
+    # Track rows with 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")
+
+    # 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]
+    if len(complete_df) == 0:
+        print("Warning: No complete rows to predict on")
+        return df
 
     # pytorch-tabular returns predictions using f"{target}_prediction" column
-    # and classification probabilities in columns ending with "_probability"
-    target = TEMPLATE_PARAMS["target_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 = []
+
+    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()}")
 
-    # Set the predictions on the DataFrame
-    df["prediction"] = predictions
+        all_ensemble_preds.append(ens_preds)
 
-    # For classification, get probabilities
+        # 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
-            df["pred_proba"] = [p.tolist() for p in probs]
+        # 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))
+    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
 
-    # All done, return the DataFrame with new columns for the predictions
     return df
 
 
@@ -336,14 +412,14 @@ if __name__ == "__main__":
     """The main function is for training the PyTorch Tabular model"""
 
     # Harness Template Parameters
-    target = TEMPLATE_PARAMS["target_column"]
+    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"]
-    validation_split = 0.2
+    hyperparameters = TEMPLATE_PARAMS["hyperparameters"]
 
     # Script arguments for input/output directories
     parser = argparse.ArgumentParser()
@@ -355,19 +431,27 @@ if __name__ == "__main__":
     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")
-    ]
+    training_files = [os.path.join(args.train, file) 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
     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
     check_dataframe(all_df, "training_df")
 
+    # Drop any rows with missing feature values
+    initial_row_count = all_df.shape[0]
+    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.")
+
     # Features/Target output
     print(f"Target: {target}")
     print(f"Features: {str(features)}")
@@ -375,125 +459,228 @@ if __name__ == "__main__":
     # Convert any features that might be categorical to 'category' type
     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}...")
         all_df, features = decompress_features(all_df, features, compressed_features)
 
-    # 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"]]
-        df_val = all_df[~all_df["training"]]
-    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}")
-
     # Determine categorical and continuous columns
-    categorical_cols = [col for col in features if df_train[col].dtype.name == "category"]
+    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}")
 
-    # Set up PyTorch Tabular configuration
-    data_config = DataConfig(
-        target=[target],
-        continuous_cols=continuous_cols,
-        categorical_cols=categorical_cols,
-    )
-
-    trainer_config = TrainerConfig(
-        auto_lr_find=True,
-        batch_size=min(1024, len(df_train) // 4),
-        max_epochs=100,
-        early_stopping="valid_loss",
-        early_stopping_patience=15,
-        checkpoints="valid_loss",
-        accelerator="auto",
-        progress_bar="none",
-        gradient_clip_val=1.0,
-    )
+    # Cast continuous columns to float
+    all_df[continuous_cols] = all_df[continuous_cols].astype("float64")
 
-    optimizer_config = OptimizerConfig()
-
-    # Choose model configuration based on model type
+    # 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
 
-    # Use CategoryEmbedding for both regression and classification tasks
-    model_config = CategoryEmbeddingModelConfig(
-        task=task,
-        layers="1024-512-512",
-        activation="ReLU",
-        learning_rate=1e-3,
-        dropout=0.1,
-        use_batch_norm=True,
-        initialization="kaiming",
-    )
+    ensemble_models = []
 
-    # Create and train the TabularModel
-    tabular_model = TabularModel(
-        data_config=data_config,
-        model_config=model_config,
-        optimizer_config=optimizer_config,
-        trainer_config=trainer_config,
+    # Set up PyTorch Tabular data configuration (shared across folds)
+    data_config = DataConfig(
+        target=[target],
+        continuous_cols=continuous_cols,
+        categorical_cols=categorical_cols,
     )
 
-    # Train the model
-    tabular_model.fit(train=df_train, validation=df_val)
+    # 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}
+
+    model_config = CategoryEmbeddingModelConfig(task=task, **model_params)
+    optimizer_config = OptimizerConfig()
 
-    # Make Predictions on the Validation Set
-    print(f"Making Predictions on Validation Set...")
-    result = tabular_model.predict(df_val, include_input_features=False)
+    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()
 
-    # pytorch-tabular returns predictions using f"{target}_prediction" column
-    # and classification probabilities in columns ending with "_probability"
-    if model_type == "classifier":
-        preds = result[f"{target}_prediction"].values
+        print(f"Fold {fold_idx + 1} complete!")
+
+    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:
-        # Regression: use the target column name
-        preds = result[f"{target}_prediction"].values
+        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 = [col for col in result.columns if col.endswith("_probability")]
-        if prob_cols:
-            probs = result[prob_cols].values
-            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)
+        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)
 
         # Decode the target and prediction labels
         y_validate = label_encoder.inverse_transform(df_val[target])
-        preds = label_encoder.inverse_transform(preds.astype(int))
+        preds_decoded = label_encoder.inverse_transform(preds.astype(int))
     else:
         y_validate = df_val[target].values
+        preds_decoded = preds
+
+    # Save predictions to S3
+    df_val = df_val.copy()
+    df_val["prediction"] = preds_decoded
+
+    # 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}")
 
-    # Save predictions to S3 (just the target, prediction, and '_probability' columns)
-    df_val["prediction"] = preds
-    output_columns = [target, "prediction"]
-    output_columns += [col for col in df_val.columns if col.endswith("_probability")]
+    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",
@@ -506,7 +693,7 @@ if __name__ == "__main__":
         label_names = label_encoder.classes_
 
         # Calculate various model performance metrics
-        scores = precision_recall_fscore_support(y_validate, preds, average=None, labels=label_names)
+        scores = precision_recall_fscore_support(y_validate, preds_decoded, average=None, labels=label_names)
 
         # Put the scores into a dataframe
         score_df = pd.DataFrame(
@@ -514,20 +701,20 @@ if __name__ == "__main__":
                 target: label_names,
                 "precision": scores[0],
                 "recall": scores[1],
-                "fscore": scores[2],
+                "f1": scores[2],
                 "support": scores[3],
             }
         )
 
-        # We need to get creative with the Classification Metrics
-        metrics = ["precision", "recall", "fscore", "support"]
+        # 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}")
 
         # Compute and output the confusion matrix
-        conf_mtx = confusion_matrix(y_validate, preds, labels=label_names)
+        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]
@@ -535,23 +722,38 @@ if __name__ == "__main__":
 
     else:
         # 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}")
-        print(f"NumRows: {len(df_val)}")
-
-    # Save the model to the standard place/name
-    tabular_model.save_model(os.path.join(args.model_dir, "tabular_model"))
+        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})")
+
     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)  # We save the original features, not the decompressed ones
+        json.dump(orig_features, fp)
 
     # Save the category mappings
     with open(os.path.join(args.model_dir, "category_mappings.json"), "w") as fp:
-        json.dump(category_mappings, fp)
+        json.dump(category_mappings, fp)