Enum anira::InferenceBackend

enum anira::InferenceBackend

Enumeration of supported neural network inference backends.

The InferenceBackend enum defines the available neural network inference engines that can be used for real-time audio processing. Each backend provides different performance characteristics, model format support, and platform compatibility.

Backend availability is determined at compile time through preprocessor macros, allowing for selective inclusion based on project requirements and dependencies. The CUSTOM backend is always available for user-defined inference implementations.

Performance and compatibility considerations:

• LIBTORCH: PyTorch models, larger memory footprint

• ONNX: Cross-platform ONNX models, optimized for CPU inference

• TFLITE: TensorFlow Lite models, optimized for mobile and embedded devices

• CUSTOM: User-defined backends for specialized inference implementations

Note

Backend availability depends on compile-time flags (USE_LIBTORCH, USE_ONNXRUNTIME, USE_TFLITE) and the presence of corresponding dependencies in the build system.

Values:

enumerator LIBTORCH

LibTorch (PyTorch C++) inference backend.

Uses the LibTorch library for running PyTorch models in C++. This backend is ideal for models trained with PyTorch. Requires the LibTorch library to be linked at build time.

Model format: .pt, .pth (PyTorch TorchScript) Platform support: Windows, Linux, macOS

enumerator ONNX

ONNX Runtime inference backend.

Uses Microsoft’s ONNX Runtime for running ONNX (Open Neural Network Exchange) format models. This backend is highly optimized for CPU inference and provides good cross-platform compatibility. Requires ONNX Runtime to be linked at build time.

Model format: .onnx Platform support: Windows, Linux, macOS, mobile platforms

enumerator TFLITE

TensorFlow Lite inference backend.

Uses Google’s TensorFlow Lite for running quantized and optimized TensorFlow models. This backend is designed for mobile and embedded devices with limited computational resources. Requires TensorFlow Lite to be linked at build time.

Model format: .tflite Platform support: Windows, Linux, macOS, Android, iOS, embedded systems

enumerator CUSTOM

Custom user-defined inference backend.

Placeholder for custom inference implementations. This backend type allows users to implement their own inference engines by extending the BackendBase class. Always available regardless of compile-time flags.

Model format: User-defined Platform support: Depends on user implementation