Inference Engine overview
Inference Engine is a neural network inference library for Unity. It lets you import trained neural network models into Unity and run them in real-time with your target device’s compute resources, such as central processing unit (CPU) or graphics processing unit (GPU).
Inference Engine supports real-time applications across all Unity-supported platforms.
The package is officially released and available to all Unity users through the Package Manager.
Tip
Prior experience with machine learning frameworks like TensorFlow or PyTorch is helpful, but not required. It can make it easier to understand how to work with models in Inference Engine.
Section | Description |
---|---|
Get started | Learn how to install Inference Engine, explore sample projects, and understand the Inference Engine workflow. |
Create a model | Create a runtime model by importing an ONNX model file or using the Inference Engine model API. |
Run a model | Create input data for a model, create an engine to run the model, and get output. |
Use Tensors | Learn how to get, set, and modify input and output data. |
Profile a model | Use Unity tools to profile the speed and performance of a model. |
Supported platforms
Inference Engine supports all Unity runtime platforms.
Performance might vary based on:
Model operators and complexity
Hardware and software platform constraints of your device
Type of engine used
For more information, refer to Models and Create an engine.
Supported model types
Inference Engine supports most models in Open Neural Network Exchange (ONNX) format with an opset version between 7 and 15. For more information, refer to Supported models and Supported ONNX operators.
Places to find pre-trained models
There are various sources to find pre-trained models, which might either be available in the ONNX format or in a format that you can convert. Examples include the following:
- Hugging Face
- Kaggle Models (Formerly TensorFlow Hub)
- PyTorch Hub
- Model Zoo
- XetData
- Meta Research
If you want to train your own models, refer to the following links: