oxionnx-core

Core types for OxiONNX -- Tensor, Graph, OpKind, Operator trait, and error types.

This crate provides the foundational data structures and abstractions used throughout the OxiONNX inference engine. It is no_std-compatible (with the std feature disabled) and has minimal dependencies.

Key Types

• Tensor -- N-dimensional tensor with f32 storage, shape, strides, and layout support (NCHW/NHWC/RowMajor).
• DType / TypedTensor -- Multi-dtype tensor support covering F32, F16, BF16, F64, I8/I16/I32/I64, U8/U16/U32/U64, and Bool.
• Graph -- Represents an ONNX computation graph as a list of Nodes with input/output names.
• OpKind -- Enum of all supported ONNX operators (147).
• Operator trait -- Stateless interface for operator implementations; receives an OpContext with resolved inputs.
• OperatorRegistry -- Maps ONNX op_type strings to Operator trait objects.
• OnnxError -- Unified error type for the engine.

Usage

[dependencies]
oxionnx-core = "0.1.0"

use oxionnx_core::{Tensor, Graph, OpKind, DType};

// Create a 2x3 tensor
let t = Tensor::new(vec![1.0, 2.0, 3.0, 4.0, 5.0, 6.0], vec![2, 3]);
assert_eq!(t.shape, vec![2, 3]);

// Layout conversion
use oxionnx_core::{nchw_to_nhwc, nhwc_to_nchw};
let img = Tensor::new(vec![0.0; 24], vec![1, 3, 2, 4]); // [N,C,H,W]
let nhwc = nchw_to_nhwc(&img).expect("layout conversion");
assert_eq!(nhwc.shape, vec![1, 2, 4, 3]); // [N,H,W,C]

Feature Flags

Part of oxionnx

A Pure Rust ONNX inference engine.

License

Apache-2.0