NeurOxyde

NeurOxyde aims to be a high performance minimal inference engine written in pure Rust. It is designed to provide a lightweight runtime for executing Open Neural Network Exchange (ONNX) models with a focus on type safety and efficiency.

Overview

NeurOxyde aims to bridge the gap between complex deep learning frameworks and embedded or edge deployment scenarios by offering a streamlined execution environment. It leverages Rust's memory safety guarantees and zero-cost abstractions to deliver reliable inference.

Key Features

• ONNX Support: Native parsing and loading of standard ONNX model files.
• Tensor Operations: Efficient n-dimensional array manipulations.
• Compute Graph: Directed acyclic graph (DAG) representation for model execution.
• Modular Design: Clear separation between model loading, graph optimization, and runtime execution.

Architecture

The project is organized into the following core modules:

• Loader: Handles the deserialization of .onnx protobuf files.
• Tensor: Provides the fundamental data structures for numerical computation.
• Ops: Implements mathematical operators (kernels) for inference.
• Graph: Manages the computational graph topology and node connectivity.
• Runtime: Orchestrates the execution flow and resource management.

Getting Started

Prerequisites

• Rust (latest stable toolchain)
• Cargo

Building

NeurOxyde uses prost for Protocol Buffers code generation. The build process automatically handles the generation of Rust bindings from the ONNX schema.

cargo build --release

Running Examples

To run the provided inference examples:

cargo run --example mnist_infer

Development

The ONNX protocol buffer definitions are located in onnx_proto/. During the build process, build.rs compiles these definitions into Rust code located in src/onnx_generated/.

To regenerate bindings after updating the .proto file:

cargo clean
cargo build