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

workbench/core/artifacts/model_core.py

@@ -21,7 +21,7 @@ from workbench.utils.aws_utils import newest_path, pull_s3_data
 from workbench.utils.s3_utils import compute_s3_object_hash
 from workbench.utils.shap_utils import shap_values_data, shap_feature_importance
 from workbench.utils.deprecated_utils import deprecated
-from workbench.utils.model_utils import proximity_model
+from workbench.utils.model_utils import published_proximity_model, get_model_hyperparameters
 
 
 class ModelType(Enum):
@@ -44,9 +44,10 @@ class ModelFramework(Enum):
     SKLEARN = "sklearn"
     XGBOOST = "xgboost"
     LIGHTGBM = "lightgbm"
-    PYTORCH_TABULAR = "pytorch_tabular"
+    PYTORCH = "pytorch"
     CHEMPROP = "chemprop"
     TRANSFORMER = "transformer"
+    META = "meta"
     UNKNOWN = "unknown"
 
 
@@ -62,7 +63,8 @@ class ModelImages:
         "inference": "py312-general-ml-inference",
         "pytorch_training": "py312-pytorch-training",
         "pytorch_inference": "py312-pytorch-inference",
-        "meta-endpoint": "py312-meta-endpoint",
+        "meta_training": "py312-meta-training",
+        "meta_inference": "py312-meta-inference",
     }
 
     @classmethod
@@ -263,21 +265,25 @@ class ModelCore(Artifact):
         else:
             self.log.important(f"No inference data found for {self.model_name}!")
 
-    def get_inference_metrics(self, capture_name: str = "latest") -> Union[pd.DataFrame, None]:
+    def get_inference_metrics(self, capture_name: str = "auto") -> Union[pd.DataFrame, None]:
         """Retrieve the inference performance metrics for this model
 
         Args:
-            capture_name (str, optional): Specific capture_name or "training" (default: "latest")
+            capture_name (str, optional): Specific capture_name (default: "auto")
         Returns:
             pd.DataFrame: DataFrame of the Model Metrics
 
         Note:
-            If a capture_name isn't specified this will try to return something reasonable
+            If a capture_name isn't specified this will try to the 'first' available metrics
         """
         # Try to get the auto_capture 'training_holdout' or the training
-        if capture_name == "latest":
-            metrics_df = self.get_inference_metrics("auto_inference")
-            return metrics_df if metrics_df is not None else self.get_inference_metrics("model_training")
+        if capture_name == "auto":
+            metric_list = self.list_inference_runs()
+            if metric_list:
+                return self.get_inference_metrics(metric_list[0])
+            else:
+                self.log.warning(f"No performance metrics found for {self.model_name}!")
+                return None
 
         # Grab the metrics captured during model training (could return None)
         if capture_name == "model_training":
@@ -299,11 +305,11 @@ class ModelCore(Artifact):
                 self.log.warning(f"Performance metrics {capture_name} not found for {self.model_name}!")
                 return None
 
-    def confusion_matrix(self, capture_name: str = "latest") -> Union[pd.DataFrame, None]:
+    def confusion_matrix(self, capture_name: str = "auto") -> Union[pd.DataFrame, None]:
         """Retrieve the confusion_matrix for this model
 
         Args:
-            capture_name (str, optional): Specific capture_name or "training" (default: "latest")
+            capture_name (str, optional): Specific capture_name or "training" (default: "auto")
         Returns:
             pd.DataFrame: DataFrame of the Confusion Matrix (might be None)
         """
@@ -315,7 +321,7 @@ class ModelCore(Artifact):
             raise ValueError(error_msg)
 
         # Grab the metrics from the Workbench Metadata (try inference first, then training)
-        if capture_name == "latest":
+        if capture_name == "auto":
             cm = self.confusion_matrix("auto_inference")
             return cm if cm is not None else self.confusion_matrix("model_training")
 
@@ -537,6 +543,17 @@ class ModelCore(Artifact):
         else:
             self.log.error(f"Model {self.model_name} is not a classifier!")
 
+    def summary(self) -> dict:
+        """Summary information about this Model
+
+        Returns:
+            dict: Dictionary of summary information about this Model
+        """
+        self.log.info("Computing Model Summary...")
+        summary = super().summary()
+        summary["hyperparameters"] = get_model_hyperparameters(self)
+        return summary
+
     def details(self) -> dict:
         """Additional Details about this Model
 
@@ -561,6 +578,7 @@ class ModelCore(Artifact):
         details["status"] = self.latest_model["ModelPackageStatus"]
         details["approval_status"] = self.latest_model.get("ModelApprovalStatus", "unknown")
         details["image"] = self.container_image().split("/")[-1]  # Shorten the image uri
+        details["hyperparameters"] = get_model_hyperparameters(self)
 
         # Grab the inference and container info
         inference_spec = self.latest_model["InferenceSpecification"]
@@ -571,16 +589,6 @@ class ModelCore(Artifact):
         details["transform_types"] = inference_spec["SupportedTransformInstanceTypes"]
         details["content_types"] = inference_spec["SupportedContentTypes"]
         details["response_types"] = inference_spec["SupportedResponseMIMETypes"]
-        details["model_metrics"] = self.get_inference_metrics()
-        if self.model_type == ModelType.CLASSIFIER:
-            details["confusion_matrix"] = self.confusion_matrix()
-            details["predictions"] = None
-        elif self.model_type in [ModelType.REGRESSOR, ModelType.UQ_REGRESSOR, ModelType.ENSEMBLE_REGRESSOR]:
-            details["confusion_matrix"] = None
-            details["predictions"] = self.get_inference_predictions()
-        else:
-            details["confusion_matrix"] = None
-            details["predictions"] = None
 
         # Grab the inference metadata
         details["inference_meta"] = self.get_inference_metadata()
@@ -869,7 +877,7 @@ class ModelCore(Artifact):
             return self.df_store.get(f"/workbench/models/{self.name}/shap_data")
         else:
             # Loop over the SHAP data and return a dict of DataFrames
-            shap_dfs = self.df_store.list_subfiles(f"/workbench/models/{self.name}/shap_data")
+            shap_dfs = self.df_store.list(f"/workbench/models/{self.name}/shap_data")
             shap_data = {}
             for df_location in shap_dfs:
                 key = df_location.split("/")[-1]
@@ -888,19 +896,19 @@ class ModelCore(Artifact):
         except (KeyError, IndexError, TypeError):
             return None
 
-    def publish_prox_model(self, prox_model_name: str = None, track_columns: list = None):
+    def publish_prox_model(self, prox_model_name: str = None, include_all_columns: bool = False):
         """Create and publish a Proximity Model for this Model
 
         Args:
             prox_model_name (str, optional): Name of the Proximity Model (if not specified, a name will be generated)
-            track_columns (list, optional): List of columns to track in the Proximity Model.
+            include_all_columns (bool): Include all DataFrame columns in results (default: False)
 
         Returns:
             Model: The published Proximity Model
         """
         if prox_model_name is None:
             prox_model_name = self.model_name + "-prox"
-        return proximity_model(self, prox_model_name, track_columns=track_columns)
+        return published_proximity_model(self, prox_model_name, include_all_columns=include_all_columns)
 
     def delete(self):
         """Delete the Model Packages and the Model Group"""

workbench/core/artifacts/parameter_store_core.py

@@ -0,0 +1,98 @@
+"""ParameterStoreCore: Manages Workbench parameters in a Cloud Based Parameter Store."""
+
+import logging
+
+# Workbench Imports
+from workbench.core.cloud_platform.aws.aws_account_clamp import AWSAccountClamp
+
+# Workbench Bridges Import
+from workbench_bridges.api import ParameterStore as BridgesParameterStore
+
+
+class ParameterStoreCore(BridgesParameterStore):
+    """ParameterStoreCore: Manages Workbench parameters in a Cloud Based Parameter Store.
+
+    Common Usage:
+        ```python
+        params = ParameterStoreCore()
+
+        # List Parameters
+        params.list()
+
+        ['/workbench/abalone_info',
+         '/workbench/my_data',
+         '/workbench/test',
+         '/workbench/pipelines/my_pipeline']
+
+        # Add Key
+        params.upsert("key", "value")
+        value = params.get("key")
+
+        # Add any data (lists, dictionaries, etc..)
+        my_data = {"key": "value", "number": 4.2, "list": [1,2,3]}
+        params.upsert("my_data", my_data)
+
+        # Retrieve data
+        return_value = params.get("my_data")
+        pprint(return_value)
+
+        {'key': 'value', 'list': [1, 2, 3], 'number': 4.2}
+
+        # Delete parameters
+        param_store.delete("my_data")
+        ```
+    """
+
+    def __init__(self):
+        """ParameterStoreCore Init Method"""
+        session = AWSAccountClamp().boto3_session
+
+        # Initialize parent with workbench config
+        super().__init__(boto3_session=session)
+        self.log = logging.getLogger("workbench")
+
+
+if __name__ == "__main__":
+    """Exercise the ParameterStoreCore Class"""
+
+    # Create a ParameterStoreCore manager
+    param_store = ParameterStoreCore()
+
+    # List the parameters
+    print("Listing Parameters...")
+    print(param_store.list())
+
+    # Add a new parameter
+    param_store.upsert("/workbench/test", "value")
+
+    # Get the parameter
+    print(f"Getting parameter 'test': {param_store.get('/workbench/test')}")
+
+    # Add a dictionary as a parameter
+    sample_dict = {"key": "str_value", "awesome_value": 4.2}
+    param_store.upsert("/workbench/my_data", sample_dict)
+
+    # Retrieve the parameter as a dictionary
+    retrieved_value = param_store.get("/workbench/my_data")
+    print("Retrieved value:", retrieved_value)
+
+    # List the parameters
+    print("Listing Parameters...")
+    print(param_store.list())
+
+    # List the parameters with a prefix
+    print("Listing Parameters with prefix '/workbench':")
+    print(param_store.list("/workbench"))
+
+    # Delete the parameters
+    param_store.delete("/workbench/test")
+    param_store.delete("/workbench/my_data")
+
+    # Out of scope tests
+    param_store.upsert("test", "value")
+    param_store.delete("test")
+
+    # Recursive delete test
+    param_store.upsert("/workbench/test/test1", "value1")
+    param_store.upsert("/workbench/test/test2", "value2")
+    param_store.delete_recursive("workbench/test/")

workbench/core/pipelines/pipeline_executor.py

@@ -123,7 +123,7 @@ class PipelineExecutor:
                 if "model" in workbench_objects and (not subset or "endpoint" in subset):
                     workbench_objects["model"].to_endpoint(**kwargs)
                     endpoint = Endpoint(kwargs["name"])
-                    endpoint.auto_inference(capture=True)
+                    endpoint.auto_inference()
 
             # Found something weird
             else:

workbench/core/transforms/features_to_model/features_to_model.py

@@ -1,6 +1,7 @@
 """FeaturesToModel: Train/Create a Model from a Feature Set"""
 
 from pathlib import Path
+from typing import Union
 from sagemaker.estimator import Estimator
 import awswrangler as wr
 from datetime import datetime, timezone
@@ -83,12 +84,17 @@ class FeaturesToModel(Transform):
         self.inference_arch = inference_arch
 
     def transform_impl(
-        self, target_column: str, description: str = None, feature_list: list = None, train_all_data=False, **kwargs
+        self,
+        target_column: Union[str, list[str]],
+        description: str = None,
+        feature_list: list = None,
+        train_all_data=False,
+        **kwargs,
     ):
         """Generic Features to Model: Note you should create a new class and inherit from
         this one to include specific logic for your Feature Set/Model
         Args:
-            target_column (str): Column name of the target variable
+            target_column (str or list[str]): Column name(s) of the target variable(s)
             description (str): Description of the model (optional)
             feature_list (list[str]): A list of columns for the features (default None, will try to guess)
             train_all_data (bool): Train on ALL (100%) of the data (default False)
@@ -105,9 +111,11 @@ class FeaturesToModel(Transform):
         s3_training_path = feature_set.create_s3_training_data()
         self.log.info(f"Created new training data {s3_training_path}...")
 
-        # Report the target column
+        # Report the target column(s)
         self.target_column = target_column
-        self.log.info(f"Target column: {self.target_column}")
+        # Normalize target_column to a list for internal use
+        target_list = [target_column] if isinstance(target_column, str) else (target_column or [])
+        self.log.info(f"Target column(s): {self.target_column}")
 
         # Did they specify a feature list?
         if feature_list:
@@ -134,7 +142,7 @@ class FeaturesToModel(Transform):
                 "is_deleted",
                 "event_time",
                 "training",
-            ] + [self.target_column]
+            ] + target_list
             feature_list = [c for c in all_columns if c not in filter_list]
 
             # AWS Feature Store has 3 user column types (String, Integral, Fractional)
@@ -157,12 +165,14 @@ class FeaturesToModel(Transform):
         self.log.important(f"Feature List for Modeling: {self.model_feature_list}")
 
         # Set up our parameters for the model script
+        # ChemProp expects target_column as a list; other templates expect a string
+        target_for_template = target_list if self.model_framework == ModelFramework.CHEMPROP else self.target_column
         template_params = {
             "model_imports": self.model_import_str,
             "model_type": self.model_type,
             "model_framework": self.model_framework,
             "model_class": self.model_class,
-            "target_column": self.target_column,
+            "target_column": target_for_template,
             "feature_list": self.model_feature_list,
             "compressed_features": feature_set.get_compressed_features(),
             "model_metrics_s3_path": self.model_training_root,
@@ -202,11 +212,13 @@ class FeaturesToModel(Transform):
         # Metric Definitions for Classification
         elif self.model_type == ModelType.CLASSIFIER:
             # We need to get creative with the Classification Metrics
+            # Note: Classification only supports single target
+            class_target = target_list[0] if target_list else self.target_column
 
             # Grab all the target column class values (class labels)
             table = feature_set.data_source.table
-            self.class_labels = feature_set.query(f'select DISTINCT {self.target_column} FROM "{table}"')[
-                self.target_column
+            self.class_labels = feature_set.query(f'select DISTINCT {class_target} FROM "{table}"')[
+                class_target
             ].to_list()
 
             # Sanity check on the targets
@@ -216,7 +228,7 @@ class FeaturesToModel(Transform):
                 raise ValueError(msg)
 
             # Dynamically create the metric definitions
-            metrics = ["precision", "recall", "f1"]
+            metrics = ["precision", "recall", "f1", "support"]
             metric_definitions = []
             for t in self.class_labels:
                 for m in metrics:
@@ -242,7 +254,7 @@ class FeaturesToModel(Transform):
         image = ModelImages.get_image_uri(self.sm_session.boto_region_name, self.training_image)
 
         # Use GPU instance for ChemProp/PyTorch, CPU for others
-        if self.model_framework in [ModelFramework.CHEMPROP, ModelFramework.PYTORCH_TABULAR]:
+        if self.model_framework in [ModelFramework.CHEMPROP, ModelFramework.PYTORCH]:
             train_instance_type = "ml.g6.xlarge"  # NVIDIA L4 GPU, ~$0.80/hr
             self.log.important(f"Using GPU instance {train_instance_type} for {self.model_framework.value}")
         else:

workbench/core/transforms/model_to_endpoint/model_to_endpoint.py

@@ -1,6 +1,7 @@
 """ModelToEndpoint: Deploy an Endpoint for a Model"""
 
 import time
+from botocore.exceptions import ClientError
 from sagemaker import ModelPackage
 from sagemaker.serializers import CSVSerializer
 from sagemaker.deserializers import CSVDeserializer
@@ -102,10 +103,21 @@ class ModelToEndpoint(Transform):
         # Is this a serverless deployment?
         serverless_config = None
         if self.serverless:
+            # For PyTorch or ChemProp we need at least 4GB of memory
+            from workbench.api import ModelFramework
+
+            self.log.info(f"Model Framework: {workbench_model.model_framework}")
+            if workbench_model.model_framework in [ModelFramework.PYTORCH, ModelFramework.CHEMPROP]:
+                if mem_size < 4096:
+                    self.log.important(
+                        f"{workbench_model.model_framework} needs at least 4GB of memory (setting to 4GB)"
+                    )
+                    mem_size = 4096
             serverless_config = ServerlessInferenceConfig(
                 memory_size_in_mb=mem_size,
                 max_concurrency=max_concurrency,
             )
+            self.log.important(f"Serverless Config: Memory={mem_size}MB, MaxConcurrency={max_concurrency}")
 
         # Configure data capture if requested (and not serverless)
         data_capture_config = None
@@ -126,16 +138,35 @@ class ModelToEndpoint(Transform):
 
         # Deploy the Endpoint
         self.log.important(f"Deploying the Endpoint {self.output_name}...")
-        model_package.deploy(
-            initial_instance_count=1,
-            instance_type=self.instance_type,
-            serverless_inference_config=serverless_config,
-            endpoint_name=self.output_name,
-            serializer=CSVSerializer(),
-            deserializer=CSVDeserializer(),
-            data_capture_config=data_capture_config,
-            tags=aws_tags,
-        )
+        try:
+            model_package.deploy(
+                initial_instance_count=1,
+                instance_type=self.instance_type,
+                serverless_inference_config=serverless_config,
+                endpoint_name=self.output_name,
+                serializer=CSVSerializer(),
+                deserializer=CSVDeserializer(),
+                data_capture_config=data_capture_config,
+                tags=aws_tags,
+            )
+        except ClientError as e:
+            # Check if this is the "endpoint config already exists" error
+            if "Cannot create already existing endpoint configuration" in str(e):
+                self.log.warning("Endpoint config already exists, deleting and retrying...")
+                self.sm_client.delete_endpoint_config(EndpointConfigName=self.output_name)
+                # Retry the deploy
+                model_package.deploy(
+                    initial_instance_count=1,
+                    instance_type=self.instance_type,
+                    serverless_inference_config=serverless_config,
+                    endpoint_name=self.output_name,
+                    serializer=CSVSerializer(),
+                    deserializer=CSVDeserializer(),
+                    data_capture_config=data_capture_config,
+                    tags=aws_tags,
+                )
+            else:
+                raise
 
     def post_transform(self, **kwargs):
         """Post-Transform: Calling onboard() for the Endpoint"""

workbench/core/transforms/pandas_transforms/pandas_to_features.py

@@ -68,6 +68,15 @@ class PandasToFeatures(Transform):
         self.output_df = input_df.copy()
         self.one_hot_columns = one_hot_columns or []
 
+        # Warn about known AWS Iceberg bug with event_time_column
+        if event_time_column is not None:
+            self.log.warning(
+                f"event_time_column='{event_time_column}' specified. Note: AWS has a known bug with "
+                "Iceberg FeatureGroups where varying event times across multiple days can cause "
+                "duplicate rows in the offline store. Setting event_time_column=None."
+            )
+            self.event_time_column = None
+
         # Now Prepare the DataFrame for its journey into an AWS FeatureGroup
         self.prep_dataframe()
 
@@ -400,7 +409,7 @@ class PandasToFeatures(Transform):
 
         # Set Hold Out Ids (if we got them during creation)
         if self.incoming_hold_out_ids:
-            self.output_feature_set.set_training_holdouts(self.id_column, self.incoming_hold_out_ids)
+            self.output_feature_set.set_training_holdouts(self.incoming_hold_out_ids)
 
     def ensure_feature_group_created(self, feature_group):
         status = feature_group.describe().get("FeatureGroupStatus")
@@ -462,7 +471,7 @@ if __name__ == "__main__":
 
     # Create my DF to Feature Set Transform (with one-hot encoding)
     df_to_features = PandasToFeatures("test_features")
-    df_to_features.set_input(data_df, id_column="id", one_hot_columns=["food"])
+    df_to_features.set_input(data_df, id_column="id", event_time_column="date", one_hot_columns=["food"])
     df_to_features.set_output_tags(["test", "small"])
     df_to_features.transform()