Metadata

Card Metadata

Both ModelCard and DataCard objects have a metadata attribute that can be used to store information about the model. If not provided, a default object is created. When registering a card, the metadata is updated with the latest information. In addition to automatically generated attributes, the metadata object can be used to store custom information about the model like descriptions.

User-Defined Attributes (Optional)

description: Description

Description object for your model

Description

Description is a simple data structure that can be used to store extra descriptive information about your model or data.

Args

Summary: Optional[str]

Summary text or pointer to a markdown file that describes the model or data

Sample Code: Optional[str]

Sample code that can be used to load and run the model or data

Notes: Optional[str]

Any additional information not captured by the other attributes

Example

from opsml import ModelCard, ModelCardMetadata
from opsml.types import Description, ModelCardMetadata

# logic for datacard or modelcard
...

modelcard = ModelCard(
  name="my_model",
  repository="my_repo",
  contact="user",
  interface=interface,
  datacard_uid=datacard.uid,
  metadata=ModelCardMetadata(description=Description(summary="my_summary.md")
  )
)

---

Docs

opsml.types.Description

Bases: BaseModel

Source code in opsml/types/extra.py

class Description(BaseModel):
    summary: Optional[str] = None
    sample_code: Optional[str] = None
    Notes: Optional[str] = None

    @field_validator("summary", mode="before")
    @classmethod
    def load_summary(cls, summary: Optional[str]) -> Optional[str]:
        if summary is None:
            return summary

        if ".md" in summary.lower():
            try:
                mkdwn_path = FileUtils.find_filepath(name=summary)
                with open(mkdwn_path, "r", encoding="utf-8") as file_:
                    summary = file_.read()

            except IndexError as idx_error:
                logger.info(f"Could not load markdown file {idx_error}")

        return summary

opsml.types.ModelCardMetadata

Bases: BaseModel

Create modelcard metadata

Parameters:

Name	Type	Description	Default
interface_type		Type of interface	required
description		Description for your model	required
data_schema		Data schema for your model	required
runcard_uid		RunCard associated with the ModelCard	required
pipelinecard_uid		Associated PipelineCard	required
auditcard_uid		Associated AuditCard	required

Source code in opsml/types/model.py

class ModelCardMetadata(BaseModel):
    """Create modelcard metadata

    Args:
        interface_type:
            Type of interface
        description:
            Description for your model
        data_schema:
            Data schema for your model
        runcard_uid:
            RunCard associated with the ModelCard
        pipelinecard_uid:
            Associated PipelineCard
        auditcard_uid:
            Associated AuditCard
    """

    interface_type: str = ""
    description: Description = Description()
    data_schema: DataSchema = DataSchema()
    runcard_uid: Optional[str] = None
    pipelinecard_uid: Optional[str] = None
    auditcard_uid: Optional[str] = None

    model_config = ConfigDict(protected_namespaces=("protect_",))

opsml.types.DataCardMetadata

Bases: BaseModel

Create a DataCard metadata

Parameters:

Name	Type	Description	Default
interface_type		Type of interface that contains data	required
data_type		Type of data	required
description		Description for your data	required
feature_map		Map of features in data (inferred when converting to pyarrow table)	required
additional_info		Dictionary of additional info to associate with data (i.e. if data is tokenized dataset, metadata could be {"vocab_size": 200})ed	required
runcard_uid		Id of RunCard that created the DataCard	required
pipelinecard_uid		Associated PipelineCard	required
auditcard_uid		Associated AuditCard	required

Source code in opsml/types/data.py

class DataCardMetadata(BaseModel):
    """Create a DataCard metadata

    Args:
        interface_type:
            Type of interface that contains data
        data_type:
            Type of data
        description:
            Description for your data
        feature_map:
            Map of features in data (inferred when converting to pyarrow table)
        additional_info:
            Dictionary of additional info to associate with data
            (i.e. if data is tokenized dataset, metadata could be {"vocab_size": 200})ed
        runcard_uid:
            Id of RunCard that created the DataCard
        pipelinecard_uid:
            Associated PipelineCard
        auditcard_uid:
            Associated AuditCard
    """

    interface_type: str = ""
    data_type: str = ""
    description: Description = Description()
    feature_map: Dict[str, Feature] = {}
    additional_info: Dict[str, Union[float, int, str]] = {}
    runcard_uid: Optional[str] = None
    pipelinecard_uid: Optional[str] = None
    auditcard_uid: Optional[str] = None

Registered Model Metadata

One of the benefits to the model registration process (especially when auto-converting to onnx) is the creation of model metadata that can be used in downstream applications to load and run models via apis or batch jobs. The example below shows sample metadata that is produced with a registered model.

Example

{
    "model_name": "regression",
    "model_class": "SklearnEstimator",
    "model_type": "LinearRegression",
    "model_interface": "SklearnModel",
    "onnx_uri": "opsml-root:/OPSML_MODEL_REGISTRY/opsml/regression/v1.4.0/onnx-model.onnx",
    "onnx_version": "1.14.1",
    "model_uri": "opsml-root:/OPSML_MODEL_REGISTRY/opsml/regression/v1.4.0/trained-model.joblib",
    "model_version": "1.4.0",
    "model_repository": "opsml",
    "sample_data_uri": "opsml-root:/OPSML_MODEL_REGISTRY/opsml/regression/v1.4.0/sample-model-data.joblib",
    "opsml_version": "2.0.0",
    "data_schema": {
        "data_type": "numpy.ndarray",
        "input_features": {
            "inputs": {
                "feature_type": "float64",
                "shape": [
                    1,
                    10
                ]
            }
        },
        "output_features": {
            "outputs": {
                "feature_type": "float64",
                "shape": [
                    1,
                    1
                ]
            }
        },
        "onnx_input_features": {
            "predict": {
                "feature_type": "tensor(float)",
                "shape": [
                    null,
                    10
                ]
            }
        },
        "onnx_output_features": {
            "variable": {
                "feature_type": "tensor(float)",
                "shape": [
                    null,
                    1
                ]
            }
        },
        "onnx_data_type": null,
        "onnx_version": "1.14.1"
    }
}

opsml.ModelMetadata

Bases: BaseModel

Model metadata associated with all registered models

Parameters:

Name	Type	Description	Default
model_name		Name of model	required
model_class		Name of model class	required
model_type		Type of model	required
model_interface		Type of interface	required
onnx_uri		URI to onnx model	required
onnx_version		Version of onnx model	required
model_uri		URI to model	required
model_version		Version of model	required
model_repository		Model repository	required
sample_data_uri		URI to sample data	required
opsml_version		Opsml version	required
data_schema		Data schema for model	required
preprocessor_uri		(only present if preprocessor is used) URI to preprocessor	required
preprocessor_name		(only present if preprocessor is used) Name of preprocessor	required
quantized_model_uri		(only present if huggingface model is quantized) URI to huggingface quantized onnx model	required
tokenizer_uri		(only present if huggingface tokenizer is used) URI to tokenizer	required
tokenizer_name		(only present if huggingface is used) Name of tokenizer	required
feature_extractor_uri		(only present if huggingface feature extractor is used) URI to feature extractor	required
feature_extractor_name		(only present if huggingface feature_extractor is used) Name of feature extractor	required

Source code in opsml/types/model.py

class ModelMetadata(BaseModel):
    """Model metadata associated with all registered models

    Args:
        model_name:
            Name of model
        model_class:
            Name of model class
        model_type:
            Type of model
        model_interface:
            Type of interface
        onnx_uri:
            URI to onnx model
        onnx_version:
            Version of onnx model
        model_uri:
            URI to model
        model_version:
            Version of model
        model_repository:
            Model repository
        sample_data_uri:
            URI to sample data
        opsml_version:
            Opsml version
        data_schema:
            Data schema for model
        preprocessor_uri: (only present if preprocessor is used)
            URI to preprocessor
        preprocessor_name: (only present if preprocessor is used)
            Name of preprocessor
        quantized_model_uri: (only present if huggingface model is quantized)
            URI to huggingface quantized onnx model
        tokenizer_uri: (only present if huggingface tokenizer is used)
            URI to tokenizer
        tokenizer_name: (only present if huggingface is used)
            Name of tokenizer
        feature_extractor_uri: (only present if huggingface feature extractor is used)
            URI to feature extractor
        feature_extractor_name: (only present if huggingface feature_extractor is used)
            Name of feature extractor
    """

    model_name: str
    model_class: str
    model_type: str
    model_interface: str
    onnx_uri: Optional[str] = None
    onnx_version: Optional[str] = None
    model_uri: str
    model_version: str
    model_repository: str
    sample_data_uri: str
    opsml_version: str = __version__
    data_schema: DataSchema

    model_config = ConfigDict(
        protected_namespaces=("protect_",),
        extra="allow",
    )