MultiModalPredictor.export_onnx

MultiModalPredictor.export_onnx(data: dict | DataFrame, path: str | None = None, batch_size: int | None = None, verbose: bool | None = False, opset_version: int | None = 16, truncate_long_and_double: bool | None = False)[source]

Export this predictor’s model to an ONNX file.

When path argument is not provided, the method would not save the model into disk. Instead, it would export the onnx model into BytesIO and return its binary as bytes.

Parameters:

  • data – Raw data used to trace and export the model. If this is None, will check if a processed batch is provided.

  • path (str, default=None) – The export path of onnx model. If path is not provided, the method would export model to memory.

  • batch_size – The batch_size of export model’s input. Normally the batch_size is a dynamic axis, so we could use a small value for faster export.

  • verbose – verbose flag in torch.onnx.export.

  • opset_version – opset_version flag in torch.onnx.export.

  • truncate_long_and_double (bool, default False) – Truncate weights provided in int64 or double (float64) to int32 and float32

Returns:

onnx_path – A string that indicates location of the exported onnx model, if path argument is provided. Otherwise, would return the onnx model as bytes.

Return type:

str or bytes