Import settings for ONNX models

You can import machine learning (ML) models in the .onnx format with Model Asset Import Settings. These settings control how Unity processes the model before it's processed at runtime.

Dynamic input shape support

Inference Engine supports ONNX models with dynamic input dimensions. This means that the model is compatible with input tensors of different shapes that match a given pattern. Dynamic dimensions in ONNX models are named, such as batch_size or sequence_length.

For example, an input with shape (batch_size, 3, 256, 256) can accept an input tensors with different batch sizes, such as:

  • (1, 3, 256, 256)
  • (2, 3, 256, 256)
  • (4, 3, 256, 256), and other values for the batch dimension.

If you know the value of a dynamic dimension in advance, for example, if the batch dimension will always be 1, you can set it to a static value. This enables Inference Engine to optimize the model for better inference speed and memory efficiency.

Configure import settings

Use the Model Asset Import Settings window to change the import settings for the model.

To set static values for dynamic dimensions, follow these steps:

  1. Open the Model Asset Import Settings from the Project window.
  2. Set a static value for any dynamic input dimensions.
  3. Select Apply.

The updated value will reflect in the Inspector for the Inference Engine model.

When you serialize the model to a .sentis file, the assigned static values are saved. However, you can’t modify input dimensions after serialization.

Dynamic input dimensions

The following table describes the properties available for dynamic input dimensions:

Property **Description
name The name of the dynamic dimension. This is automatically populated from the ONNX model, such as batch_size or sequence_length. You can't modify this field.
value The value assigned to a dynamic input dimension. A value of -1 keeps the dimension dynamic, while any non-negative value (≥ 0) sets it to a static size.

For more information on the shape of your model inputs, refer to Model inputs.

Additional resources