dataeval.utils.onnx.to_encoding_model¶

dataeval.utils.onnx.to_encoding_model(model_path: str | pathlib.Path, output_path: None = None, embedding_layer: str | None = None) → tuple[bytes, str]¶

dataeval.utils.onnx.to_encoding_model(model_path: str | pathlib.Path, output_path: str | pathlib.Path, embedding_layer: str | None = None) → tuple[str, str]

Modify an ONNX model to output embeddings from an intermediate layer.

This function modifies a classification model to expose an intermediate layer as an additional output, allowing extraction of embeddings instead of (or in addition to) classification logits.

Parameters:¶

model_path : str or Path¶

Path to the original ONNX model file.

output_path : str, Path, or None, default None¶

Path where the modified model will be saved. If None, returns serialized model bytes instead of writing to disk.

embedding_layer : str or None, default None¶

Name of the layer to expose as output. If None, automatically detected using find_embedding_layer().

Returns:¶

If output_path is None: (model_bytes, layer_name) If output_path is provided: (output_path, layer_name)

Return type:¶

tuple[bytes, str] or tuple[str, str]

Raises:¶

• ImportError – If the onnx package is not installed.

• ValueError – If the embedding layer cannot be identified (when embedding_layer is None).