teradataml 20.0.0.2__py3-none-any.whl → 20.0.0.3__py3-none-any.whl

teradataml/opensource/sklearn/_sklearn_wrapper.py

@@ -19,7 +19,6 @@ from collections import OrderedDict, defaultdict
 from importlib import import_module
 
 import base64
-import functools
 import json
 import numpy
 import os
@@ -28,7 +27,7 @@ import time
 import inspect
 import warnings
 import json
-import random
+import math
 import pandas as pd
 from teradatasqlalchemy import BLOB, CLOB, FLOAT, TIMESTAMP, VARCHAR, INTEGER
 import pandas.api.types as pt
@@ -41,9 +40,9 @@ from teradataml.context.context import _get_current_databasename, get_connection
 from teradataml.dbutils.filemgr import install_file, remove_file
 from teradataml.utils.utils import execute_sql
 from teradataml.options.configure import configure
-from teradataml.opensource.sklearn._wrapper_utils import _validate_fit_run, _generate_new_name,\
+from teradataml.opensource._wrapper_utils import _validate_fit_run, _generate_new_name,\
     _validate_opensource_func_args, _derive_df_and_required_columns, _validate_df_query_type
-from teradataml.opensource.sklearn.constants import OpenSourcePackage, _OSML_MODELS_PRIMARY_INDEX,\
+from teradataml.opensource.constants import OpenSourcePackage, _OSML_MODELS_PRIMARY_INDEX,\
     _OSML_MODELS_TABLE_NAME, _OSML_MODELS_TABLE_COLUMNS_TYPE_DICT, OpensourceModels,\
     _OSML_ADDITIONAL_COLUMN_TYPES
 from teradataml.common.messagecodes import MessageCodes
@@ -53,7 +52,6 @@ from teradataml.dbutils.dbutils import _create_table, set_session_param
 from teradataml.utils.validators import _Validators
 from teradataml.dataframe.dataframe import DataFrame
 from teradataml.dataframe.dataframe_utils import DataFrameUtils
-from teradataml.scriptmgmt.lls_utils import create_env, get_env
 from teradataml.common.garbagecollector import GarbageCollector
 from teradataml.common.constants import TeradataConstants
 
@@ -70,6 +68,9 @@ _file_installed = False
 
 class _GenericObjectWrapper:
     def __init__(self) -> None:
+        if not get_connection():
+            raise TeradataMlException(Messages.get_message(MessageCodes.INVALID_CONTEXT_CONNECTION),
+                                      MessageCodes.INVALID_CONTEXT_CONNECTION)
         self._db_name = _get_current_databasename()
 
         self._scripts_path = os.path.join(_TDML_DIRECTORY, "data", "scripts", "sklearn")
@@ -215,6 +216,7 @@ class _GenericObjectWrapper:
             raise TeradataMlException(
                 f"Script file '{file_name}' failed to remove in Vantage."
             )
+
     def _get_data_col_types_and_partition_col_indices_and_types(self, data, partition_columns,
                                                                 idx_delim=",",
                                                                 types_delim="--"):
@@ -264,7 +266,7 @@ class _GenericObjectWrapper:
                 args_str += f" {strr}"
         return args_str
 
-    def extract_sklearn_obj(self, n_unique_partitions = 1, n_partition_cols = 1):
+    def _extract_model_objs(self, n_unique_partitions=1, n_partition_cols=1):
         """
         Internal function to extract sklearn object from the model(s) depending on the number of
         partitions. When it is only one model, it is directly used as sklearn object (modelObj).
@@ -297,13 +299,256 @@ class _GenericObjectWrapper:
 
         warnings.filterwarnings("default")
 
+    def _validate_existence_of_partition_columns(self, partition_columns, all_columns, arg_names_for_dfs):
+        """
+        Validate if columns in "partition_columns" argument are present in any of the given
+        dataframes.
+        """
+        invalid_part_cols = [c for c in partition_columns if c not in all_columns]
+
+        if invalid_part_cols:
+            raise ValueError(Messages.get_message(MessageCodes.INVALID_PARTITIONING_COLS,
+                                                  ", ".join(invalid_part_cols),
+                                                  "', '".join(arg_names_for_dfs))
+                                                  )
+
+    def _prepare_data_args_string(self, kwargs):
+        """
+        Get column indices and types of each data related arguments in the format:
+        "{<arg_name>-<comma separated indices>-<comma separated types>}--
+         {<arg_name>-<comma separated indices>-<comma separated types>}"
+        """
+        data_args_str = []
+        for arg_name in list(self._data_args.keys()):
+            # Remove DataFrame arguments from kwargs, which will be passed to Script.
+            kwargs.pop(arg_name)
+
+            # Get column indices and their types for each dataframe from parent dataframe.
+            _, partition_indices_str, partition_types_str, _ = \
+                self._get_data_col_types_and_partition_col_indices_and_types(self._tdml_df,
+                                                                   self._data_args[arg_name].columns,
+                                                                   idx_delim=",",
+                                                                   types_delim=",")
+        
+            # Format "<arg_name>-<comma separated indices>-<comma separated types>"            
+            data_args_str.append(f"{arg_name}-{partition_indices_str}-{partition_types_str}")
+        
+        # Format "{<arg_name>-<comma separated indices>-<comma separated types>}--
+        #    {<arg_name>-<comma separated indices>-<comma separated types>}"
+        return "--".join(data_args_str)
+
+    def _prepare_and_install_file(self, replace_dict):
+        """
+        Prepare function script file from template file and install it in Vantage.
+        Takes the dictionary with keys as strings to be replaced in script and values as
+        strings which should be added in place of keys.
+        """
+
+        with open(os.path.join(self._scripts_path, self._template_file)) as fp:
+            script_data = fp.read()
+        
+        for old, new in replace_dict.items():
+            script_data = script_data.replace(old, new)
+
+        self._script_file_local = os.path.join(self._tdml_tmp_dir, self._script_file_name)
+
+        with open(self._script_file_local, "w") as fp:
+            fp.write(script_data)
+        
+        self._install_script_file(file_identifier=self._script_file_name.split(".")[0],
+                                  file_name=self._script_file_name,
+                                  file_location=self._tdml_tmp_dir)
+
+    def _get_dataframe_related_args_and_their_columns(self, kwargs):
+        """
+        Get dataframe related arguments and return all their column names from kwargs.
+        """
+        __data_columns = []
+        __data_args_dict = OrderedDict()
+
+        # Separate dataframe related arguments and their column names from actual kwargs.
+        for k, v in kwargs.items():
+            if isinstance(v, DataFrame):
+                # All dataframes should be select of parent dataframe.
+                _validate_df_query_type(v, "select", k)
+
+                # Save all columns in dataframe related arguments.
+                __data_columns.extend(v.columns)
+
+                __data_args_dict[k] = v
+        
+        return __data_args_dict, __data_columns
+
+    def _process_data_for_funcs_returning_objects(self, kwargs):
+        """
+        Internal function to process all arguments and assign self._data_args, self._tdml_df
+        and return 
+        1. dictionary of elements (needed to replace in the script template file)
+        2. partition columns list.
+        """
+        partition_cols = self._get_columns_as_list(kwargs.get("partition_columns", None))
+        if partition_cols:
+            kwargs.pop("partition_columns")
+
+        self._data_args, __data_columns = self._get_dataframe_related_args_and_their_columns(kwargs)
+
+        arg_names_for_dfs = list(self._data_args.keys())
+
+        # Get common parent dataframe from all dataframes.
+        self._tdml_df =  DataFrameUtils()._get_common_parent_df_from_dataframes(list(self._data_args.values()))
+
+        self._tdml_df = self._tdml_df.select(__data_columns + partition_cols)
+
+        self._validate_existence_of_partition_columns(partition_cols, self._tdml_df.columns, arg_names_for_dfs)
+
+        self._tdml_df, partition_cols = self._get_data_and_data_partition_columns(self._tdml_df,
+                                                                                   __data_columns,
+                                                                                   [],
+                                                                                   partition_cols
+                                                                                   )
+
+        # Prepare string of data arguments with name, indices where columns of that argument resides
+        # and types of each of the column.
+        data_args_str = self._prepare_data_args_string(kwargs)
+
+        # Get indices of partition_columns and types of all columns.
+        data_column_types_str, partition_indices_str, _, partition_cols = \
+            self._get_data_col_types_and_partition_col_indices_and_types(self._tdml_df,
+                                                                         partition_cols,
+                                                                         types_delim=None,
+                                                                         idx_delim=None)
+
+        replace_dict = {"<partition_cols_indices>": str(partition_indices_str),
+                        "<types_of_data_cols>": str(data_column_types_str),
+                        "<data_args_info_str>": f"'{data_args_str}'"}
+
+        return replace_dict, partition_cols
+
+    def _validate_equality_of_partition_values(self, fit_values, trans_values):
+        """
+        Internal function to compare the partition values in fit() and predict() are same.
+        """
+        if len(fit_values) != len(trans_values):
+            return False
+
+        for val in fit_values:
+            if not all([val in trans_values]):
+                return False
+
+        return True
+
+    def _get_non_data_related_args_from_kwargs(self, kwargs):
+        """
+        Get all non-data related arguments from kwargs.
+        """
+        non_data_related_args = {}
+        for k, v in kwargs.items():
+            if not isinstance(v, DataFrame):
+                non_data_related_args[k] = v
+        non_data_related_args.pop("partition_columns", None)
+        return non_data_related_args
+
+    def _read_from_template_and_write_dict_to_file(self, template_file, replace_dict,
+                                                   output_script_file_name=None):
+        """
+        Read template file, replace the keys with values and write to new file.
+        """
+        with open(os.path.join(self._scripts_path, template_file)) as fp:
+            script_data = fp.read()
+        
+        for old, new in replace_dict.items():
+            script_data = script_data.replace(old, new)
+
+        if output_script_file_name is None:
+            output_script_file_name = self._script_file_name
+        file_path = os.path.join(self._tdml_tmp_dir, output_script_file_name)
+        with open(file_path, "w") as fp:
+            fp.write(script_data)
+
+    def _generate_script_file_from_template_file(self, kwargs, template_file, func_name,
+                                                 output_script_file_name=None):
+        """
+        Internal function to generate script file from template file. It just adds the non-data
+        related arguments to the template file and writes the contents to new file, so that these
+        arguments are available in the script file for running this function "func_name".
+        """
+        # Take out all non-data related arguments to write to template file.
+        non_data_related_args = self._get_non_data_related_args_from_kwargs(kwargs)
+
+        # Read template file and write the contents to new file with non-data related arguments.
+        template_f = os.path.join(self._scripts_path, template_file)
+        with open(template_f, "r") as f:
+            template = f.read()
+
+        if output_script_file_name is None:
+            output_script_file_name = self._script_file_name
+        file_path = os.path.join(self._tdml_tmp_dir, output_script_file_name)
+        with open(file_path, "w") as f:
+            f.write("import json\n")
+            f.write(f"params = json.loads('{json.dumps(non_data_related_args)}')\n")
+            f.write(template)
+
+        kwargs["file_name"] = output_script_file_name
+        kwargs["name"] = func_name
+
+    def _remove_data_related_args_from_kwargs(self, kwargs):
+        """
+        Internal function to remove data related arguments from kwargs.
+        """
+        kwargs.pop("data", None)
+        kwargs.pop("feature_columns", None)
+        kwargs.pop("group_columns", None)
+        kwargs.pop("partition_columns", None)
+        kwargs.pop("label_columns", None)
+
+    def _convert_pos_args_to_kwargs_for_function(self, pos_args, kwargs, func_name):
+        """
+        Internal function to convert positional arguments to keyword arguments.
+        """
+        fn = getattr(getattr(import_module(self.module_name), self.class_name), func_name)
+        kwargs.update(zip(fn.__code__.co_varnames[1:], pos_args))
+
+    def _install_model_and_script_files(self, file_name, file_location):
+        """
+        Internal function to install model and script files to Vantage.
+        """
+        self._install_initial_model_file()
+        self._install_script_file(file_identifier=file_name.split(".")[0],
+                                  file_name=file_name,
+                                  is_binary=False,
+                                  file_location=file_location)
+
+    def _assign_fit_variables_after_execution(self, data, partition_columns, label_columns):
+        """
+        Internal function to assign fit related variables.
+        """
+        # Extract sklearn object(s) from the depending on the number of unique partitioning values.
+        self._extract_model_objs(n_unique_partitions=len(self._fit_partition_unique_values),
+                                 n_partition_cols=len(partition_columns))
+
+        # Need this label columns types in prediction.
+        self._fit_label_columns_types = []
+        self._fit_label_columns_python_types = []
+
+        for l_c in label_columns:
+            column_data = data._td_column_names_and_sqlalchemy_types[l_c.lower()]
+            self._fit_label_columns_types.append(column_data)
+            self._fit_label_columns_python_types.append(column_data.python_type.__name__)
+
+        # If the model is trained a second time after the object creation,
+        # or if set_params() is called after the first model training,
+        # this flag will reset to False. So that for subsequent predict/score
+        # operations, the newly trained model will be installed.
+        if self._is_trained_model_installed:
+            self._is_trained_model_installed = False
+
 
 class _OpenSourceObjectWrapper(_GenericObjectWrapper):
     # This has to be set for every package which subclasses this class.
     OPENSOURCE_PACKAGE_NAME = None
 
     def __init__(self, model=None, module_name=None, class_name=None, pos_args=None, kwargs=None):
-        if not model and not module_name and not class_name:
+        if model is None and not module_name and not class_name:
             raise TeradataMlException(Messages.get_message(MessageCodes.EITHER_THIS_OR_THAT_ARGUMENT, "model",
                                                            "module_name and class_name"),
                                       MessageCodes.EITHER_THIS_OR_THAT_ARGUMENT)
@@ -319,24 +564,224 @@ class _OpenSourceObjectWrapper(_GenericObjectWrapper):
         self.pos_args = pos_args if pos_args is not None else tuple()
 
         self._fit_label_columns_types = None
+        self._fit_label_columns_python_types = None
         self._table_name_prefix = None
 
         self._is_default_partition_value_fit = True # False when the user provides partition columns.
         self._fit_partition_colums_non_default = None
         self._is_default_partition_value_predict = True # False when the user provides partition columns.
 
-    def _validate_equality_of_partition_values(self, fit_values, trans_values):
+    def __repr__(self):
+        if self._is_default_partition_value_fit:
+            # Single model use case.
+            return self.modelObj.__repr__()
+
+        pd.set_option("display.expand_frame_repr", None)
+        pd.set_option("display.max_colwidth", None)
+        opt = self.modelObj.__repr__()
+        pd.reset_option("display.expand_frame_repr")
+        pd.reset_option("display.max_colwidth")
+        return opt
+
+    def _initialize_object(self):
         """
-        Internal function to compare the partition values in fit() and predict() are same.
+        Internal function to initialize sklearn object from module name and class name.
         """
-        if len(fit_values) != len(trans_values):
-            return False
+        # Needed when writing imported modules to generated file. TODO: Remove later.
+        imported_args = {} 
+        # If there are any objects of class `_SkLearnObjectWrapper`, it is modified to
+        # corresponding sklearn object.
+        _partition_column_names = None
+        if "partition_columns" in self.kwargs:
+            self._fit_partition_colums_non_default = self.kwargs["partition_columns"]
+            self._is_default_partition_value_fit = False
+            _partition_column_names = self._fit_partition_colums_non_default
 
-        for val in fit_values:
-            if not all([val in trans_values]):
-                return False
 
-        return True
+        new_sklearn_pos_args = self.modify_args(None, self.pos_args, imported_args)
+        new_sklearn_kwargs = self.modify_args(None, self.kwargs, imported_args)
+
+        # Create model object from new positional and keyword arguments.
+        class_obj = getattr(import_module(self.module_name), self.class_name)
+        if new_sklearn_pos_args:
+            self.modelObj = class_obj(*new_sklearn_pos_args, **new_sklearn_kwargs)
+        else:
+            self.modelObj = class_obj(**new_sklearn_kwargs)
+
+        # All arguments are moved to kwargs and kept pos_args empty.
+        # Might help in set_params() bug fix.
+        self.pos_args = tuple()
+        _arguments = self.modelObj.__dict__
+
+        if hasattr(self.modelObj, "get_params"):
+            # Update kwargs that are both in modelObj and get_params() as there are
+            # some classes which return other internals variables also. 
+            # Hence, filtering them using get_params().
+            for k, v in _arguments.items():
+                if type(v).__name__ in ["function", "generator"]:
+                    # TODO: ELE-6351: Skipping adding functions and generators to kwargs as these
+                    #       are not supported yet due to pickling issue.
+                    continue
+                if self.get_params():
+                    if k in self.get_params():
+                        self.kwargs[k] = v
+                else:
+                    _model_init_arguments = None
+                    try:
+                        _model_init_arguments = self.modelObj.__init__.__code__.co_varnames
+                    except AttributeError:
+                        pass
+                    if _model_init_arguments:
+                        self.kwargs = dict((k, v) for k, v in _arguments.items() if k in _model_init_arguments)
+                    else:
+                        self.kwargs = _arguments
+        else:
+            # Model selection classes will not have `get_params`, in which case modelObj's __dict__
+            # is saved as kwargs.
+            self.kwargs = _arguments
+
+        if _partition_column_names:
+            self.kwargs["partition_columns"] = _partition_column_names
+
+    def _initialize_variables(self, table_name_prefix):
+        """
+        Internal function to initialize variables used in this class.
+        """
+        self.feature_names_in_ = None
+        self._table_name_prefix = table_name_prefix
+        self._model_file_name_prefix = _generate_new_name(type="file")
+        self.model_file_paths_local = set()
+
+        self._fit_execution_time = None
+        self._fit_predict_execution_time = None
+        self._partial_fit_execution_time = None
+        self._predict_execution_time = None
+        self._transform_execution_time = None
+        self._score_execution_time = None
+
+        # Set to partition columns when training is done with partition columns.
+        self._fit_partition_colums_non_default = None
+
+        self._is_model_installed = False
+        self._fit_partition_unique_values = [[self._default_data_partition_value]]
+
+    def _get_returning_df(self, script_df, partition_column, returns):
+        """
+        Internal function to return the teradataml Dataframe except
+        partition_column.
+        """
+        if self._is_default_partition_value_fit:
+            # For single model case, partition column is internally generated
+            # and no point in returning it to the user.
+
+            # Extract columns from return types.
+            returning_cols = [col[0] for col in returns[len(partition_column):]]
+            return script_df.select(returning_cols)
+        return script_df
+
+    def modify_args(self, fp1, arg, imported_args):
+        """
+        Internal function to recursively (if "arg" is list/tuple/dict) check if any sklearn object
+        of opensourceML is present in the argument "arg" and modify it to corresponding sklearn
+        object.
+        This function can also be used to write import statements to file (if "fp1" is not
+        None). Update "imported_args" dictionary with imported module and class name to avoid
+        importing same module and class again when writing to file. This is useful when we want to
+        generate script from template file.
+        Pass None to "fp1" if we don't want to write to file and just modify opensourceML sklearn
+        object to corresponding sklearn object.
+        """
+        if isinstance(arg, type(self)):
+            imported_tuple = (arg.module_name, arg.class_name)
+            already_imported = imported_args.get(imported_tuple, False)
+            if not already_imported:
+                imported_args[imported_tuple] = True
+                if fp1:
+                    fp1.write(f"from {arg.module_name} import {arg.class_name}\n")
+                self.modify_args(fp1, arg.pos_args, imported_args)
+                self.modify_args(fp1, arg.kwargs, imported_args)
+            return arg.modelObj
+        elif isinstance(arg, list):
+            return [self.modify_args(fp1, val, imported_args) for val in arg]
+        elif isinstance(arg, tuple):
+            return tuple([self.modify_args(fp1, val, imported_args) for val in arg])
+        elif type(arg).__name__ == "generator":
+            # Raising exception as generator object can't be pickled.
+            # TODO: ELE-6351 - Find ways to pickle generator object later.
+            raise ValueError("Generator type/iterator is not supported for any argument. "\
+                             "Support will be added later.")
+        elif type(arg).__name__ == "function":
+            # Raising exception as functions/lambda functions can't be pickled.
+            # TODO: ELE-6351 - Find ways to pickle functions later.
+            raise ValueError("Functions are not supported for any argument. "\
+                             "Support will be added later.")
+        elif isinstance(arg, dict):
+            return dict(
+                (
+                    self.modify_args(fp1, k, imported_args),
+                    self.modify_args(fp1, v, imported_args),
+                )
+                for k, v in arg.items() if k != "partition_columns"
+            )
+        # elif arg == "partition_columns":
+
+        else:
+            return arg
+
+    def _install_initial_model_file(self, use_dummy_initial_file=False):
+        """
+        If model file(s) is/are not installed in Vantage, then install it/them.
+        """
+        if isinstance(self.modelObj, pd.DataFrame):
+            # Get list of unique partition values and corresponding model object as dict.
+            partition_values_model_dict = {}
+            obj_list = self.modelObj.values.tolist()
+            for lst in obj_list:
+                partition_values_model_dict[tuple(lst[:len(self._fit_partition_colums_non_default)])] = \
+                    lst[len(self._fit_partition_colums_non_default)]
+
+        for partition in self._fit_partition_unique_values:
+            # Create a new file with file name with partition values and
+            # dump sklearn object into it. Finally install the file to Vantage.
+            partition_join = "_".join([str(x) for x in partition])
+            file_name = f"{self._model_file_name_prefix}_{partition_join}"
+            # Replace '-' with '_' as '-' can't be present in file identifier.
+            # Needed this replace because partition_columns can be negative.
+            file_name = file_name.replace("-", "_")
+            full_file_name = os.path.join(self._tdml_tmp_dir, file_name)
+            with open(full_file_name, "wb+") as fp:
+                # Write sklearn object to file.
+                if isinstance(self.modelObj, pd.DataFrame):
+                    # If multiple models, then write the model corresponding to the partition value.
+                    fp.write(pickle.dumps(partition_values_model_dict[tuple(partition)]))
+                else:
+                    if use_dummy_initial_file:
+                        fp.write(pickle.dumps("abc"))
+                    else:
+                        fp.write(pickle.dumps(self.modelObj))
+            self.model_file_paths_local.add(file_name)
+
+            self._install_script_file(file_identifier=file_name,
+                                      file_name=file_name,
+                                      is_binary=True,
+                                      file_location=self._tdml_tmp_dir)
+
+            if self._is_lake_system:
+                # Need to pass env_name along with file_name for cleaning up the files in env.
+                obj = f"{self._env.env_name}::{file_name}"
+                if installed_model_files[obj] == 0:
+                    # Add to GC for the first time the model file (along with env name) is encountered.
+                    installed_model_files[obj] = 1
+                    GarbageCollector._add_to_garbagecollector(object_name=obj,
+                                                object_type=TeradataConstants.TERADATA_APPLY)
+            else:
+                if installed_model_files[file_name] == 0:
+                    # Add to GC for the first time the model file is encountered.
+                    installed_model_files[file_name] = 1
+                    GarbageCollector._add_to_garbagecollector(object_name=file_name,
+                                                object_type=TeradataConstants.TERADATA_SCRIPT)
+
+            self._is_model_installed = True
 
     def _validate_unique_partition_values(self, data, partition_columns):
         """
@@ -361,25 +806,61 @@ class _OpenSourceObjectWrapper(_GenericObjectWrapper):
 
         if not self._validate_equality_of_partition_values(fit_unique_values, trans_unique_values):
             raise TeradataMlException(
-                Messages.get_message(MessageCodes.PARTITION_VALUES_NOT_MATCHING),
+                Messages.get_message(MessageCodes.PARTITION_VALUES_NOT_MATCHING, "training", "test"),
                 MessageCodes.PARTITION_VALUES_NOT_MATCHING
             )
 
     def fit(self, **kwargs):
         pass
 
+    def _convert_arguments_to_modelObj(self, args, idx_multi_model=None):
+        """
+        Internal function to convert all OpensourceML related objects in arguments to
+        underlying model objects.
+        """
+        if isinstance(args, dict):
+            new_args = args.copy() # To avoid updating 
+            for k, v in new_args.items():
+                if isinstance(v, type(self)):
+                    if idx_multi_model is not None:
+                        # single model. This argument is set only when modelObj is single model.
+                        new_args[k] = v.modelObj
+                    else:
+                        # multi-model. Get appropriate model from modelObj.
+                        new_args[k] = v.modelObj.iloc[idx_multi_model]["model"]
+                else:
+                    new_args[k] = v
+            return new_args
+        
+        # If args is tuple, convert all elements to underlying model object.
+        elif isinstance(args, tuple):
+            new_args = tuple()
+            for arg in args:
+                if isinstance(arg, type(self)):
+                    if idx_multi_model is None:
+                        # single model. This argument is set only when modelObj is single model.
+                        new_args += (arg.modelObj,)
+                    else:
+                        # multi-model. Get appropriate model from modelObj.
+                        new_args += (arg.modelObj.iloc[idx_multi_model]["model"],)
+                else:
+                    new_args += (arg,)
+            return new_args
+        return args
+
     def __get_obj_attributes_multi_model(self, name):
         """
         Internal function to get attributes of all sklearn model objects when multiple models are
         generated by fit.
         """
 
-        def __generate_model_object(model_obj_value):
+        def __generate_model_object(model_obj_value, init_model_obj):
             """
             Internal function to generate _SkLearnWrapperObject model object from model_obj_value.
             """
             # Create _SkLearnObjectWrapper object from opensource model object.
-            model_obj = self.__class__(model=first_atrribute_instance)
+            model_obj = self.__class__(model=init_model_obj)
+
             model_obj.modelObj = model_obj_value
             model_obj._is_model_installed = True
 
@@ -396,13 +877,34 @@ class _OpenSourceObjectWrapper(_GenericObjectWrapper):
             multi_models = self.modelObj.copy()
             for i in range(multi_models.shape[0]):
                 curr_model = multi_models.iloc[i]["model"]
-                multi_models.at[i, "model"] = getattr(curr_model, name)(*c, **kwargs)
+                partition_values = multi_models.iloc[i][0:len(self._fit_partition_colums_non_default)].to_list()
+                partition_values = "_".join([str(x) for x in partition_values])
+                if self.module_name == "lightgbm.basic" and self.class_name == "Booster" and name == "save_model":
+                    # filename is first argument.
+                    kwargs1 = kwargs.copy()
+                    c1 = c
+
+                    if len(c) > 0:
+                        c1 = list(c1)
+                        c1[0] = f"{c1[0]}_{partition_values}"
+                        c1 = tuple(c1)
+                    if len(kwargs) > 0 and kwargs.get("filename", None):
+                        kwargs1["filename"] = f"{kwargs1['filename']}_{partition_values}"
+                    
+                    multi_models.at[i, "model"] = getattr(curr_model, name)(*self._convert_arguments_to_modelObj(c1, i),
+                                                                            **self._convert_arguments_to_modelObj(kwargs1, i))
+                else:
+                    multi_models.at[i, "model"] = getattr(curr_model, name)(*self._convert_arguments_to_modelObj(c, i),
+                                                                            **self._convert_arguments_to_modelObj(kwargs, i))
             
-            first_function_instance = multi_models.at[0, "model"]
-            if self.__class__._validate_model_supportability(first_function_instance):
-                return __generate_model_object(multi_models)
+            first_function_value = multi_models.at[0, "model"]
+            if self.__class__._validate_model_supportability(first_function_value):
+                return __generate_model_object(multi_models, init_model_obj=first_function_value)
 
-            return multi_models.rename(columns={"model": name})
+            multi_models = multi_models.rename(columns={"model": name})
+
+            # Select only partition columns and the attribute column.
+            return multi_models[self._fit_partition_colums_non_default + [name]]
 
         # Assuming that self.modelObj will have at least 1 row.
 
@@ -420,15 +922,15 @@ class _OpenSourceObjectWrapper(_GenericObjectWrapper):
             output_attributes.at[i, "model"] = getattr(model, name)
 
         if self.__class__._validate_model_supportability(first_atrribute_instance):
-            return __generate_model_object(output_attributes)
+            return __generate_model_object(output_attributes, init_model_obj=first_atrribute_instance)
 
         return output_attributes.rename(columns={"model": name})
 
     def __getattr__(self, name):
-        # This just run attributes (functions and properties) from sklearn object.
+        # This just run attributes (functions and properties) from opensource (sklearn/lightgbm) objects.
         def __sklearn_method_invoker(*c, **kwargs):
-            # sklearn model is returned from the function call. Create _SkLearnObjectWrapper object.
-            model_obj = attribute_instance(*c, **kwargs)
+            # Opensource model is returned from the function call. Create _OpensourceObjectWrapper object.
+            model_obj = attribute_instance(*self._convert_arguments_to_modelObj(c), **self._convert_arguments_to_modelObj(kwargs))
             if self.__class__._validate_model_supportability(model_obj):
                 model_obj = self.__class__(model=model_obj)
                 model_obj._is_model_installed = True # Trained model is returned by function call.
@@ -636,234 +1138,63 @@ class _OpenSourceObjectWrapper(_GenericObjectWrapper):
         EXAMPLES:
             >>> from teradataml import td_sklearn
             >>> model = td_sklearn.LinearRegression(normalize=True)
-            >>> model
-            LinearRegression(normalize=True)
-
-            # Example 1: Deploy the model held by interface object to Vantage.
-            >>> lin_reg = model.deploy("linreg_model_ver_2")
-            Model is saved.
-            >>> lin_reg
-            LinearRegression(normalize=True)
-
-            # Example 2: Deploy the model held by interface object to Vantage with the name same
-            #            as that of model that already existed in Vantage.
-            >>> lin_reg = model.deploy("linreg_model_ver_2", replace_if_exists=True)
-            Model is deleted.
-            Model is saved.
-            >>> lin_reg
-            LinearRegression(normalize=True)
-        """
-
-        # Install model file into Vantage, if not installed.
-        self._install_initial_model_file()
-
-        self._save_model(model_name, replace_if_exists)
-        return self
-
-
-class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
-
-    OPENSOURCE_PACKAGE_NAME = OpenSourcePackage.SKLEARN
-
-    def __init__(self, model=None, module_name=None, class_name=None, pos_args=None, kwargs=None):
-        super().__init__(model=model, module_name=module_name, class_name=class_name,
-                         pos_args=pos_args, kwargs=kwargs)
-
-        self._initialize_variables()
-        if model:
-            self.modelObj = model
-            self.module_name = model.__module__.split("._")[0]
-            self.class_name = model.__class__.__name__
-            # __dict__ gets all the arguments as dictionary including default ones and positional
-            # args.
-            self.kwargs = model.__dict__
-            self.pos_args = tuple() # Kept empty as all are moved to kwargs.
-        else:
-            self._initialize_object()
-
-    def __repr__(self):
-        if self._is_default_partition_value_fit:
-            # Single model use case.
-            return self.modelObj.__repr__()
-
-        pd.set_option("display.expand_frame_repr", None)
-        pd.set_option("display.max_colwidth", None)
-        opt = self.modelObj.__repr__()
-        pd.reset_option("display.expand_frame_repr")
-        pd.reset_option("display.max_colwidth")
-        return opt
-
-    def _validate_args_and_get_data(self, X=None, y=None, groups=None, kwargs={},
-                                    skip_either_or_that=False):
-        """
-        Internal function to validate arguments passed to exposed opensource APIs and return
-        parent DataFrame, feature columns, label columns, group columns, data partition columns.
-        """
-        _validate_opensource_func_args(X=X, y=y, groups=groups,
-                                       fit_partition_cols=self._fit_partition_colums_non_default,
-                                       kwargs=kwargs,
-                                       skip_either_or_that=skip_either_or_that)
-        return _derive_df_and_required_columns(X=X, y=y, groups=groups, kwargs=kwargs,
-                                        fit_partition_cols=self._fit_partition_colums_non_default)
-
-    def _initialize_object(self):
-        """
-        Internal function to initialize sklearn object from module name and class name.
-        """
-        # Needed when writing imported modules to generated file. TODO: Remove later.
-        imported_args = {} 
-        # If there are any objects of class `_SkLearnObjectWrapper`, it is modified to
-        # corresponding sklearn object.
-        new_sklearn_pos_args = self.modify_args(None, self.pos_args, imported_args)
-        new_sklearn_kwargs = self.modify_args(None, self.kwargs, imported_args)
-
-        # Create model object from new positional and keyword arguments.
-        class_obj = getattr(import_module(self.module_name), self.class_name)
-        if new_sklearn_pos_args:
-            self.modelObj = class_obj(*new_sklearn_pos_args, **new_sklearn_kwargs)
-        else:
-            self.modelObj = class_obj(**new_sklearn_kwargs)
-
-        # All arguments are moved to kwargs and kept pos_args empty.
-        # Might help in set_params() bug fix.
-        self.pos_args = tuple()
-        _arguments = self.modelObj.__dict__
-
-        if hasattr(self.modelObj, "get_params"):
-            # Update kwargs that are both in modelObj and get_params() as there are
-            # some classes which return other internals variables also. 
-            # Hence, filtering them using get_params().
-            for k, v in _arguments.items():
-                if type(v).__name__ in ["function", "generator"]:
-                    # TODO: ELE-6351: Skipping adding functions and generators to kwargs as these
-                    #       are not supported yet due to pickling issue.
-                    continue
-                if k in self.get_params():
-                    self.kwargs[k] = v
-        else:
-            # Model selection classes will not have `get_params`, in which case modelObj's __dict__
-            # is saved as kwargs.
-            self.kwargs = _arguments
-
-    def _initialize_variables(self):
-        """
-        Internal function to initialize variables used in this class.
-        """
-        self.feature_names_in_ = None
-        self._table_name_prefix = "td_sklearn_"
-        self._model_file_name_prefix = _generate_new_name(type="file")
-        self.model_file_paths_local = set()
-
-        self._fit_execution_time = None
-        self._fit_predict_execution_time = None
-        self._partial_fit_execution_time = None
-        self._predict_execution_time = None
-        self._transform_execution_time = None
-        self._score_execution_time = None
-
-        # Set to partition columns when training is done with partition columns.
-        self._fit_partition_colums_non_default = None
-
-        self._is_model_installed = False
-        self._fit_partition_unique_values = [[self._default_data_partition_value]]
-
-    def modify_args(self, fp1, arg, imported_args):
-        """
-        Internal function to recursively (if "arg" is list/tuple/dict) check if any sklearn object
-        of opensourceML is present in the argument "arg" and modify it to corresponding sklearn
-        object.
-        This function can also be used to write import statements to file (if "fp1" is not
-        None). Update "imported_args" dictionary with imported module and class name to avoid
-        importing same module and class again when writing to file. This is useful when we want to
-        generate script from template file.
-        Pass None to "fp1" if we don't want to write to file and just modify opensourceML sklearn
-        object to corresponding sklearn object.
-        """
-        if isinstance(arg, type(self)):
-            imported_tuple = (arg.module_name, arg.class_name)
-            already_imported = imported_args.get(imported_tuple, False)
-            if not already_imported:
-                imported_args[imported_tuple] = True
-                if fp1:
-                    fp1.write(f"from {arg.module_name} import {arg.class_name}\n")
-                self.modify_args(fp1, arg.pos_args, imported_args)
-                self.modify_args(fp1, arg.kwargs, imported_args)
-            return arg.modelObj
-        elif isinstance(arg, list):
-            return [self.modify_args(fp1, val, imported_args) for val in arg]
-        elif isinstance(arg, tuple):
-            return tuple([self.modify_args(fp1, val, imported_args) for val in arg])
-        elif type(arg).__name__ == "generator":
-            # Raising exception as generator object can't be pickled.
-            # TODO: ELE-6351 - Find ways to pickle generator object later.
-            raise ValueError("Generator type/iterator is not supported for any argument. "\
-                             "Support will be added later.")
-        elif type(arg).__name__ == "function":
-            # Raising exception as functions/lambda functions can't be pickled.
-            # TODO: ELE-6351 - Find ways to pickle functions later.
-            raise ValueError("Functions are not supported for any argument. "\
-                             "Support will be added later.")
-        elif isinstance(arg, dict):
-            return dict(
-                (
-                    self.modify_args(fp1, k, imported_args),
-                    self.modify_args(fp1, v, imported_args),
-                )
-                for k, v in arg.items()
-            )
-        else:
-            return arg
+            >>> model
+            LinearRegression(normalize=True)
 
-    def _install_initial_model_file(self):
-        """
-        If model file(s) is/are not installed in Vantage, then install it/them.
+            # Example 1: Deploy the model held by interface object to Vantage.
+            >>> lin_reg = model.deploy("linreg_model_ver_2")
+            Model is saved.
+            >>> lin_reg
+            LinearRegression(normalize=True)
+
+            # Example 2: Deploy the model held by interface object to Vantage with the name same
+            #            as that of model that already existed in Vantage.
+            >>> lin_reg = model.deploy("linreg_model_ver_2", replace_if_exists=True)
+            Model is deleted.
+            Model is saved.
+            >>> lin_reg
+            LinearRegression(normalize=True)
         """
-        if isinstance(self.modelObj, pd.DataFrame):
-            # Get list of unique partition values and corresponding model object as dict.
-            partition_values_model_dict = {}
-            obj_list = self.modelObj.values.tolist()
-            for lst in obj_list:
-                partition_values_model_dict[tuple(lst[:len(lst)-1])] = lst[len(lst)-1]
 
-        for partition in self._fit_partition_unique_values:
-            # Create a new file with file name with partition values and
-            # dump sklearn object into it. Finally install the file to Vantage.
-            partition_join = "_".join([str(x) for x in partition])
-            file_name = f"{self._model_file_name_prefix}_{partition_join}"
-            # Replace '-' with '_' as '-' can't be present in file identifier.
-            # Needed this replace because partition_columns can be negative.
-            file_name = file_name.replace("-", "_")
-            full_file_name = os.path.join(self._tdml_tmp_dir, file_name)
-            with open(full_file_name, "wb+") as fp:
-                # Write sklearn object to file.
-                if isinstance(self.modelObj, pd.DataFrame):
-                    # If multiple models, then write the model corresponding to the partition value.
-                    fp.write(pickle.dumps(partition_values_model_dict[tuple(partition)]))
-                else:
-                    fp.write(pickle.dumps(self.modelObj))
-            self.model_file_paths_local.add(file_name)
+        # Install model file into Vantage, if not installed.
+        self._install_initial_model_file()
 
-            self._install_script_file(file_identifier=file_name,
-                                      file_name=file_name,
-                                      is_binary=True,
-                                      file_location=self._tdml_tmp_dir)
+        self._save_model(model_name, replace_if_exists)
+        return self
 
-            if self._is_lake_system:
-                # Need to pass env_name along with file_name for cleaning up the files in env.
-                obj = f"{self._env.env_name}::{file_name}"
-                if installed_model_files[obj] == 0:
-                    # Add to GC for the first time the model file (along with env name) is encountered.
-                    installed_model_files[obj] = 1
-                    GarbageCollector._add_to_garbagecollector(object_name=obj,
-                                                object_type=TeradataConstants.TERADATA_APPLY)
-            else:
-                if installed_model_files[file_name] == 0:
-                    # Add to GC for the first time the model file is encountered.
-                    installed_model_files[file_name] = 1
-                    GarbageCollector._add_to_garbagecollector(object_name=file_name,
-                                                object_type=TeradataConstants.TERADATA_SCRIPT)
 
-            self._is_model_installed = True
+class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
+
+    OPENSOURCE_PACKAGE_NAME = OpenSourcePackage.SKLEARN
+
+    def __init__(self, model=None, module_name=None, class_name=None, pos_args=None, kwargs=None):
+        super().__init__(model=model, module_name=module_name, class_name=class_name,
+                         pos_args=pos_args, kwargs=kwargs)
+
+        self._initialize_variables(table_name_prefix="td_sklearn_")
+        if model is not None:
+            self.modelObj = model
+            self.module_name = model.__module__.split("._")[0]
+            self.class_name = model.__class__.__name__
+            # __dict__ gets all the arguments as dictionary including default ones and positional
+            # args.
+            self.kwargs = model.__dict__
+            self.pos_args = tuple() # Kept empty as all are moved to kwargs.
+        else:
+            self._initialize_object()
+
+    def _validate_args_and_get_data(self, X=None, y=None, groups=None, kwargs={},
+                                    skip_either_or_that=False):
+        """
+        Internal function to validate arguments passed to exposed opensource APIs and return
+        parent DataFrame, feature columns, label columns, group columns, data partition columns.
+        """
+        _validate_opensource_func_args(X=X, y=y, groups=groups,
+                                       fit_partition_cols=self._fit_partition_colums_non_default,
+                                       kwargs=kwargs,
+                                       skip_either_or_that=skip_either_or_that)
+        return _derive_df_and_required_columns(X=X, y=y, groups=groups, kwargs=kwargs,
+                                        fit_partition_cols=self._fit_partition_colums_non_default)
 
     def _run_fit_related_functions(self,
                                    data,
@@ -871,7 +1202,8 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
                                    label_columns,
                                    partition_columns,
                                    func,
-                                   classes=None):
+                                   classes=None,
+                                   file_name="sklearn_fit.py"):
         """
         Internal function to run fit() and partial_fit() functions.
         """
@@ -886,8 +1218,6 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
         return_types = [(col, data._td_column_names_and_sqlalchemy_types[col.lower()]) 
                         for col in new_partition_columns] + [("model", model_type)]
 
-        file_name = "sklearn_fit.py"
-
         if classes:
             class_type = type(classes[0]).__name__
             classes = "--".join([str(x) for x in classes])
@@ -913,20 +1243,7 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
         self._model_data = self._run_script(data, script_command, new_partition_columns,
                                             return_types)
 
-        # Extract sklearn object(s) from the depending on the number of unique partitioning values.
-        self.extract_sklearn_obj(n_unique_partitions=len(self._fit_partition_unique_values),
-                                 n_partition_cols=len(new_partition_columns))
-
-        # Need this label columns types in prediction.
-        self._fit_label_columns_types = [data._td_column_names_and_sqlalchemy_types[l_c.lower()]
-                                         for l_c in label_columns]
-
-        # If the model is trained a second time after the object creation,
-        # or if set_params() is called after the first model training,
-        # this flag will reset to False. So that for subsequent predict/score
-        # operations, the newly trained model will be installed.
-        if self._is_trained_model_installed:
-            self._is_trained_model_installed = False
+        self._assign_fit_variables_after_execution(data, new_partition_columns, label_columns)
 
     def partial_fit(self, X=None, y=None, classes=None, **kwargs):
         """
@@ -974,11 +1291,19 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
             self._is_default_partition_value_fit = False
             self._fit_partition_colums_non_default = partition_columns
 
-        self._run_fit_related_functions(data,
-                                        feature_columns,
-                                        label_columns,
-                                        partition_columns,
-                                        inspect.stack()[0][3])
+        file_name = kwargs.pop("file_name", None)
+        func_name = kwargs.pop("name", "fit")
+
+        args = {"data": data,
+                "feature_columns": feature_columns,
+                "label_columns": label_columns,
+                "partition_columns": partition_columns,
+                "func": func_name}
+        
+        if file_name is not None:
+            args["file_name"] = file_name
+
+        self._run_fit_related_functions(**args)
 
         self._fit_execution_time = time.time() - st_time
 
@@ -1043,10 +1368,130 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
 
         return super().__getattr__(name)
 
+    def _special_handling_multimodel_(self, data, feature_columns, label_columns, partition_columns,
+                                      func_name, **kwargs):
+        """
+        Internal function to handle multi model case for transform function for functions 
+        ["SelectFpr", "SelectFdr", "SelectFwe", "SelectFromModel", "RFECV"] of feature_selection module
+        and "Birch" of cluster module.
+        These functions generate multiple models and when transform is applied to each model, it generates
+        output with different number of columns.
+        """
+        skl_objs_dict = {}
+        no_of_unique_partitions = len(self._fit_partition_unique_values)
+        no_of_partitioning_cols = len(self._fit_partition_unique_values[0])
+
+        # Run on 10 rows of data individually using corresponding scikit-learn objects based on paritition value
+        # and get the maximum number of columns and their types.
+        for i in range(no_of_unique_partitions):
+            skl_objs_dict[tuple(self.modelObj.iloc[i, :no_of_partitioning_cols])] = self.modelObj.iloc[i]["model"]
+        
+
+        data = data.select(feature_columns + label_columns + partition_columns)
+        ten_row_data = data.head(10).get_values()
+        X = numpy.array(ten_row_data)
+
+        # For multi-model case, model in one AMP can give more number of columns than other AMPs.
+        # Returns clause can't contain different number of columns in different AMPs. Hence, taking
+        # maximum number of columns and their types from all models.
+        max_no_of_columns = 0
+        max_col_names = []
+        max_col_types = []
+
+        def _get_input_row_without_nans(row):
+            """
+            `inverse_transform` should not contain NaNs. Hence, removing NaNs from the row.
+            """
+            X1 = []
+            for _, v in enumerate(row):
+                if isinstance(v, type(None)) or isinstance(v, str) or not math.isnan(v) or self.module_name == "sklearn.impute":
+                    # Add to list when:
+                    #  - v is None or
+                    #   - v is string or
+                    #   - v is not nan or
+                    #   - if module is impute (which transforms nan values) even though v is nan.
+                    X1.append(v)
+                else:
+                    # skip nan values.
+                    pass
+            return X1
+
+        for i in range(X.shape[0]):
+            # Run `transform` or `inverse_transform` on each row with corresponding scikit-learn model object.
+            partition_values = tuple(X[i, -no_of_partitioning_cols:])
+            skl_obj = skl_objs_dict[partition_values]
+
+            X1 = X[i, :-no_of_partitioning_cols]
+            # Since Nans/NULLs are added in transform for last columns where some models generated
+            # less number of columns, removing Nans/NULLs from the input row for inverse_transform
+            # using function _get_input_row_without_nans().
+            X1 = numpy.array([_get_input_row_without_nans(X1)])
+
+            trans_opt = getattr(skl_obj, func_name)(X1, **kwargs)
+
+            no_of_columns = 1
+
+            if trans_opt.shape == (X1.shape[0],):
+                trans_opt = trans_opt.reshape(X1.shape[0], 1)
+            
+            if isinstance(trans_opt[0], numpy.ndarray) \
+                    or isinstance(trans_opt[0], list) \
+                    or isinstance(trans_opt[0], tuple):
+                no_of_columns = len(trans_opt[0])
+            
+            col_names = [f"{self.class_name.lower()}_{func_name}_{(i + 1)}" for i in range(no_of_columns)]
+
+            # Get new column sqlalchemy types for pandas df columns of transform output.
+            opt_pd = pd.DataFrame(trans_opt)
+
+            # Get output column types for each column in pandas df from the output of transform
+            # type functions.
+            types = {}
+            for idx in range(no_of_columns):
+                col = list(opt_pd.columns)[idx]
+
+                # Only one row in trans_opt.
+                if isinstance(trans_opt[0], numpy.ndarray) or isinstance(trans_opt[0], tuple) or isinstance(trans_opt[0], list):
+                    type_ = type(trans_opt[0][idx])
+                else:
+                    # only one value in the output.
+                    type_ = type(trans_opt[0])
+
+                # If type of the output value (trans_opt) is None, then use `str` as type since
+                # pandas astype() does not accept None type.
+                if type_ is type(None):
+                    type_ = str
+
+                # numpy integer columns with nan values can't be typecasted using pd.astype() to int64.
+                # It raises error like "Cannot convert non-finite values (NA or inf) to integer: 
+                #                       Error while type casting for column '2'"
+                # Hence, using pd.Int64Dtype() for integer columns with nan values.
+                types[col] = type_ if type_ not in [int, numpy.int64] else pd.Int64Dtype()
+
+            # Without this, all columns will be of object type and gets converted to VARCHAR in Vantage.
+            opt_pd = opt_pd.astype(types)
+
+            # If the datatype is not specified then check if the datatype is datetime64 and timezone is present then map it to
+            # TIMESTAMP(timezone=True) else map it according to default value.
+            col_types = [TIMESTAMP(timezone=True)
+                        if pt.is_datetime64_ns_dtype(opt_pd.dtypes[key]) and (opt_pd[col_name].dt.tz is not None)
+                        else _get_sqlalchemy_mapping(str(opt_pd.dtypes[key]))
+                        for key, col_name in enumerate(list(opt_pd.columns))]
+
+            # Different models in multi model case can generate different number of output columns for example in
+            # SelectFpr. Hence, taking the model which generates maximum number of columns.
+            if no_of_columns > max_no_of_columns:
+                max_no_of_columns = no_of_columns
+                max_col_names = col_names
+                max_col_types = col_types
+
+        return [(c_name, c_type) for c_name, c_type in zip(max_col_names, max_col_types)]
+
     def _get_return_columns_for_function_(self,
                                           data,
                                           feature_columns,
                                           label_columns,
+                                          partition_columns,
                                           func_name,
                                           kwargs):
         """
@@ -1060,7 +1505,8 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
             return [(f"{self.class_name.lower()}_{func_name}_{(i + 1)}",
                      data._td_column_names_and_sqlalchemy_types[col.lower()])
                     for i, col in enumerate(label_columns)]
-        if func_name == "predict":
+
+        if func_name == "predict" and self.OPENSOURCE_PACKAGE_NAME == OpenSourcePackage.SKLEARN:
             """
             Return predict columns using either label_columns (if provided) or 
             self._fit_label_columns_types (if the function is trained using label columns).
@@ -1075,8 +1521,6 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
                 return [(f"{self.class_name.lower()}_{func_name}_{(i + 1)}", col_type)
                         for i, col_type in enumerate(self._fit_label_columns_types)]
 
-        data = data.select(feature_columns + label_columns)
-
         ## If function is not `fit_predict`:
         #   then take one row of transform/other functions to execute in client
         #   to get number of columns in return clause and their Vantage types.
@@ -1090,8 +1534,20 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
             skl_obj = self.modelObj
         else:
             # Multi model case.
+            if (func_name in ["transform", "inverse_transform"] and \
+                self.class_name in ["SelectFpr", "SelectFdr", "SelectFwe", "SelectFromModel", "RFECV", "Birch"]) or \
+                (self.module_name == "lightgbm.sklearn" and self.class_name == "LGBMClassifier"):
+                # Special handling for multi model case for transform function as these classes
+                # generate transform output with different number of columns for each model.
+                # Hence, need to add Nulls/Nans to columns which are not present in the transform output of
+                # some models.
+                return self._special_handling_multimodel_(data, feature_columns, label_columns,
+                                                          partition_columns, func_name, **kwargs)
+
             skl_obj = self.modelObj.iloc[0]["model"]
 
+        data = data.select(feature_columns + label_columns)
+
         ten_row_data = data.head(10).get_values()
         X = numpy.array(ten_row_data)
         if label_columns:
@@ -1200,7 +1656,7 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
         return [(c_name, c_type) for c_name, c_type in zip(col_names, col_types)]
 
     @_validate_fit_run
-    def _run_function_needing_all_rows(self, X=None, y=None, **kwargs):
+    def _run_function_needing_all_rows(self, X=None, y=None, file_name="sklearn_score.py", **kwargs):
         """
         Internal function to run functions like score, aic, bic which needs all rows and return
         one floating number as result.
@@ -1223,8 +1679,6 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
                                                                                 label_columns,
                                                                                 partition_columns)
 
-        file_name = "sklearn_score.py"
-
         script_file_path = f"{file_name}" if self._is_lake_system \
             else f"./{self._db_name}/{file_name}"
 
@@ -1260,7 +1714,7 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
         return opt
 
     @_validate_fit_run
-    def _transform(self, X=None, y=None, **kwargs):
+    def _transform(self, X=None, y=None, file_name="sklearn_transform.py", **kwargs):
         """
         Internal function to run predict/transform and similar functions, which returns
         multiple columns. This function will return data row along with the generated
@@ -1283,18 +1737,7 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
                                                                                 partition_columns)
 
         # Since kwargs are passed to transform, removing additional unrelated arguments from kwargs.
-        if "data" in kwargs:
-            kwargs.pop("data")
-        if "feature_columns" in kwargs:
-            kwargs.pop("feature_columns")
-        if "group_columns" in kwargs:
-            kwargs.pop("group_columns")
-        if "partition_columns" in kwargs:
-            kwargs.pop("partition_columns")
-        if "label_columns" in kwargs:
-            kwargs.pop("label_columns")
-
-        file_name = "sklearn_transform.py"
+        self._remove_data_related_args_from_kwargs(kwargs)
 
         script_file_path = f"{file_name}" if self._is_lake_system \
             else f"./{self._db_name}/{file_name}"
@@ -1304,24 +1747,36 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
 
         self._validate_unique_partition_values(data, new_partition_columns)
 
-        py_exc = UtilFuncs._get_python_execution_path()
-        script_command = f"{py_exc} {script_file_path} {func_name} {len(feature_columns)} "\
-            f"{len(label_columns)} {partition_indices_str} {data_column_types_str} "\
-            f"{self._model_file_name_prefix} {self._is_lake_system}"
+        return_columns_python_types = None
+        if self._fit_label_columns_python_types:
+            return_columns_python_types = '--'.join(self._fit_label_columns_python_types)
 
         # Returning feature columns also along with transformed columns because we don't know the
         # mapping of feature columns to the transformed columns.
-        return_types = [(col, data._td_column_names_and_sqlalchemy_types[col.lower()])
-                        for col in (new_partition_columns + feature_columns)]
+        ## 'correct_covariance()' returns the (n_features, n_features)
+        if func_name == "correct_covariance":
+            return_types = [(col, data._td_column_names_and_sqlalchemy_types[col.lower()])
+                            for col in new_partition_columns]
+        else:
+            return_types = [(col, data._td_column_names_and_sqlalchemy_types[col.lower()])
+                            for col in (new_partition_columns + feature_columns)]
         if func_name in ["predict", "decision_function"] and label_columns:
             return_types += [(col, data._td_column_names_and_sqlalchemy_types[col.lower()])
                              for col in label_columns]
 
-        return_types += self._get_return_columns_for_function_(data,
-                                                               feature_columns,
-                                                               label_columns,
-                                                               func_name,
-                                                               kwargs)
+        output_cols_types = self._get_return_columns_for_function_(data,
+                                                                   feature_columns,
+                                                                   label_columns,
+                                                                   new_partition_columns,
+                                                                   func_name,
+                                                                   kwargs)
+        return_types += output_cols_types
+
+        py_exc = UtilFuncs._get_python_execution_path()
+        script_command = f"{py_exc} {script_file_path} {func_name} {len(feature_columns)} "\
+            f"{len(label_columns)} {partition_indices_str} {data_column_types_str} "\
+            f"{self._model_file_name_prefix} {len(output_cols_types)} {self._is_lake_system} " \
+            f"{return_columns_python_types}"
 
         # Checking the trained model installation. If not installed,
         # install it and set flag to True.
@@ -1363,6 +1818,7 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
             return_types += self._get_return_columns_for_function_(data,
                                                                    feature_columns,
                                                                    label_columns,
+                                                                   new_partition_columns,
                                                                    func_name,
                                                                    {})
         else:
@@ -1448,14 +1904,10 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
                                              skip_either_or_that=True)
 
         # Remove the kwargs data.
-        input_data = kwargs.pop("data", None)
-        partition_cols = kwargs.pop("partition_columns", None)
-        feature_cols = kwargs.pop("feature_columns", None)
-        label_cols = kwargs.pop("label_columns", None)
+        self._remove_data_related_args_from_kwargs(kwargs)
 
         if partition_columns:
             # kwargs are passed to kneighbors function. So, removing them from kwargs.
-            kwargs.pop("partition_columns")
             self._is_default_partition_value_fit = False
 
         # Generating new partition column name.
@@ -1640,161 +2092,69 @@ class _SkLearnObjectWrapper(_OpenSourceObjectWrapper):
 
         return opt
 
-    def _get_returning_df(self, script_df, partition_column, returns):
-        """
-        Internal function to return the teradataml Dataframe except
-        partition_column.
-        """
-        if self._is_default_partition_value_fit:
-            # For single model case, partition column is internally generated
-            # and no point in returning it to the user.
-
-            # Extract columns from return types.
-            returning_cols = [col[0] for col in returns[len(partition_column):]]
-            return script_df.select(returning_cols)
-        return script_df
 
-
-class _SKLearnFunctionWrapper(_GenericObjectWrapper):
-    def __init__(self, module_name, func_name):
+class _FunctionWrapper(_GenericObjectWrapper):
+    def __init__(self, module_name, func_name, file_type, template_file):
         super().__init__()
-        self.__module_name = module_name
-        self.__func_name = func_name
-        self.__params = None
-        self.__data_args = OrderedDict()
-        self._model_file_name = _generate_new_name(type="file_function", extension="py")
+        self._module_name = module_name
+        self._func_name = func_name
+        self._params = None
+        self._data_args = OrderedDict()
+        self._template_file = template_file
+        self._script_file_name = _generate_new_name(type=file_type, extension="py")
 
     def __call__(self, **kwargs):
         """
         Run the function with all the arguments passed from `td_sklearn.<function_name>` function.
         """
-        __data_columns = []
-
-        partition_cols = self._get_columns_as_list(kwargs.get("partition_columns", None))
-        if partition_cols:
-            kwargs.pop("partition_columns")
-
-        # Separate dataframe related arguments and their column names from actual kwargs.
-        for k, v in kwargs.items():
-            if isinstance(v, DataFrame):
-                # All dataframes should be select of parent dataframe.
-                _validate_df_query_type(v, "select", k)
-
-                # Save all columns in dataframe related arguments.
-                __data_columns.extend(v.columns)
-
-                self.__data_args[k] = v
-
-
-        # Get common parent dataframe from all dataframes.
-        self.__tdml_df =  DataFrameUtils()._get_common_parent_df_from_dataframes(list(self.__data_args.values()))
+        replace_dict, partition_cols = self._process_data_for_funcs_returning_objects(kwargs)
 
-        self._validate_existence_of_partition_columns(partition_cols, self.__tdml_df.columns)
-
-        self.__tdml_df = self.__tdml_df.select(__data_columns + partition_cols)
-
-        self.__tdml_df, partition_cols = self._get_data_and_data_partition_columns(self.__tdml_df,
-                                                                                   __data_columns,
-                                                                                   [],
-                                                                                   partition_cols
-                                                                                   )
-
-        # Prepare string of data arguments with name, indices where columns of that argument resides
-        # and types of each of the column.
-        data_args_str = self._prepare_data_args_string(kwargs)
-
-        self.__params = kwargs
-
-        # Get indices of partition_columns and types of all columns.
-        data_column_types_str, partition_indices_str, _, partition_cols = \
-            self._get_data_col_types_and_partition_col_indices_and_types(self.__tdml_df, partition_cols)
-
-        script_file_path = f"{self._model_file_name}" if self._is_lake_system \
-            else f"./{self._db_name}/{self._model_file_name}"
+        script_file_path = f"{self._script_file_name}" if self._is_lake_system \
+            else f"./{self._db_name}/{self._script_file_name}"
 
         model_file_prefix = None
         if self._is_lake_system:
-            model_file_prefix = self._model_file_name.replace(".py", "")
+            model_file_prefix = self._script_file_name.replace(".py", "")
 
         py_exc = UtilFuncs._get_python_execution_path()
-        script_command = (f"{py_exc} {script_file_path} {partition_indices_str} "\
-                          f"{data_column_types_str} {data_args_str} {self._is_lake_system}"\
-                          f" {model_file_prefix}")
+        script_command = f"{py_exc} {script_file_path} {model_file_prefix} {self._is_lake_system}"
 
         model_type = BLOB() if self._is_lake_system else CLOB()
-        return_types = [(col, self.__tdml_df._td_column_names_and_sqlalchemy_types[col.lower()])
-                        for col in partition_cols] + [(self.__func_name, model_type)]
-
-        # Generate new file in .teradataml directory and install it to Vantage.
-        self._prepare_and_install_file()
-
-        self._model_data = self._run_script(self.__tdml_df, script_command, partition_cols, return_types)
-        self._model_data._index_label = None
 
-        fit_partition_unique_values = self.__tdml_df.drop_duplicate(partition_cols).get_values()
+        return_types = [(col, self._tdml_df._td_column_names_and_sqlalchemy_types[col.lower()]) 
+                        for col in partition_cols] + [(self._func_name, model_type)]
 
-        self.extract_sklearn_obj(n_unique_partitions=len(fit_partition_unique_values),
-                                 n_partition_cols=len(partition_cols))
-
-        # File cleanup after processing.
-        os.remove(self._model_file_local)
-        self._remove_script_file(self._model_file_name)
+        replace_dict.update({"<module_name>": self._module_name,
+                             "<func_name>": self._func_name,
+                             "<params>": json.dumps(kwargs)})
 
-        return self.modelObj
+        # Generate new file in .teradataml directory and install it to Vantage.
+        self._prepare_and_install_file(replace_dict=replace_dict)
 
-    def _prepare_data_args_string(self, kwargs):
-        """
-        Get column indices and types of each data related arguments in the format:
-        "{<arg_name>-<comma separated indices>-<comma separated types>}--
-         {<arg_name>-<comma separated indices>-<comma separated types>}"
-        """
-        data_args_str = []
-        for arg_name in list(self.__data_args.keys()):
-            # Remove DataFrame arguments from kwargs, which will be passed to Script.
-            kwargs.pop(arg_name)
+        try:
+            self._model_data = self._run_script(self._tdml_df, script_command, partition_cols, return_types)
+            self._model_data._index_label = None
 
-            # Get column indices and their types for each dataframe from parent dataframe.
-            _, partition_indices_str, partition_types_str, _ = \
-                self._get_data_col_types_and_partition_col_indices_and_types(self.__tdml_df,
-                                                                   self.__data_args[arg_name].columns,
-                                                                   idx_delim=",",
-                                                                   types_delim=",")
-        
-            # Format "<arg_name>-<comma separated indices>-<comma separated types>"            
-            data_args_str.append(f"{arg_name}-{partition_indices_str}-{partition_types_str}")
-        
-        # Format "{<arg_name>-<comma separated indices>-<comma separated types>}--
-        #    {<arg_name>-<comma separated indices>-<comma separated types>}"
-        return "--".join(data_args_str)
+            fit_partition_unique_values = self._tdml_df.drop_duplicate(partition_cols).get_values()
 
-    def _validate_existence_of_partition_columns(self, partition_columns, all_columns):
-        """
-        Validate if columns in "partition_columns" argument are present in any of the given
-        dataframes.
-        """
-        invalid_part_cols = [c for c in partition_columns if c not in all_columns]
+            self._extract_model_objs(n_unique_partitions=len(fit_partition_unique_values),
+                                     n_partition_cols=len(partition_cols))
 
-        if invalid_part_cols:
-            raise ValueError(Messages.get_message(MessageCodes.INVALID_PARTITIONING_COLS,
-                                                  ", ".join(invalid_part_cols),
-                                                  "', '".join(list(self.__data_args.keys())))
-                                                  )
+        except Exception as ex:
+            # File cleanup if script execution fails or unable to fetch modelObj.
+            os.remove(self._script_file_local)
+            self._remove_script_file(self._script_file_name)
+            raise
 
-    def _prepare_and_install_file(self):
-        """
-        Prepare function script file from template file and install it in Vantage.
-        """
-        with open(os.path.join(self._scripts_path, "sklearn_function.template")) as fp:
-            script_data = fp.read()
-        script_data = script_data.replace("<module_name>",self.__module_name).\
-            replace("<func_name>",self.__func_name).replace("<params>", json.dumps(self.__params))
+        # File cleanup after processing.
+        os.remove(self._script_file_local)
+        self._remove_script_file(self._script_file_name)
 
-        self._model_file_local = os.path.join(self._tdml_tmp_dir, self._model_file_name)
+        return self.modelObj
 
-        with open(self._model_file_local, "w") as fp:
-            fp.write(script_data)
-        
-        self._install_script_file(file_identifier=self._model_file_name.split(".")[0],
-                                  file_name=self._model_file_name,
-                                  file_location=self._tdml_tmp_dir)
 
+class _SKLearnFunctionWrapper(_FunctionWrapper):
+    def __init__(self, module_name, func_name):
+        file_type = "file_fn_sklearn"
+        template_file = "sklearn_function.template"
+        super().__init__(module_name, func_name, file_type=file_type, template_file=template_file)