Export and convert a file to ONNX

Inference Engine imports files in the Open Neural Network Exchange (ONNX) format or the .sentis serialized format. If your model is in a different format, convert it to ONNX before you import it into Unity.

The following sections describe how to export and convert models to the ONNX format.

Export an ONNX file from a machine learning framework

Most machine learning frameworks let you to export models in ONNX format.

To export files in ONNX format from common machine learning frameworks, refer to the following documentation:

Note

To ensure compatibility with Inference Engine, set the ONNX opset version to 15 during export. For more information about ONNX compatibility, refer to import a model file.

Convert TensorFlow files to ONNX

TensorFlow uses two primary file types for saving models: SavedModel and Checkpoints.

The following sections explain each file type and how to convert them to the ONNX format.

Model files

TensorFlow saves models in SavedModel files, which contain a complete TensorFlow program, including trained parameters and computation. SavedModels have the .pb file extension. For more information on SavedModels, refer to Using the SavedModel format (TensorFlow documentation).

To generate an ONNX file from a SavedModel, use the tf2onnx tool. This is a command line tool that works best with full path names.

Checkpoints

Checkpoints (TensorFlow documentation) contain only the parameters of the model.

Checkpoints in TensorFlow can consist up of two file formats:

  • A file to store the graph, with the extension .ckpt.meta.
  • A file to store the weights, with the extension .ckpt.

If you have both the graph and weight file types, use the tf2onnx tool to create an ONNX file.

If you only have the .ckpt file, find the Python code that constructs the model and loads in the weights. After that, proceed to export the model to ONNX.

Convert PyTorch files to ONNX

The following sections explain how to convert PyTorch files to the ONNX format.

Model files

PyTorch model files usually have the .pt file extension.

To export a model file to ONNX, refer to the links in the following instructions:

  1. Load the model in Python.
  2. Export the model as an ONNX file. When you export your model, it's recommended to use Opset 15 or higher.

If your .pt file doesn't contain the model graph, find the Python code that constructs the model and loads in the weights. After that, export the model to ONNX.

Checkpoints

You can create Checkpoints in PyTorch to save the state of your model at any instance of time. Checkpoint files are usually denoted with the .tar or .pth extension.

To convert a checkpoint file to ONNX, find the Python code which constructs the model and loads in the weights. After that, proceed to export the model to ONNX.

Additional resources