https://github.com/sonos/tract
Somehow this has never been suggested (or I didn't find it 
). A neural network inference library, written purely in Rust for models in ONNX, NNEF and TF formats. Very extensive support for many different operators and comparable speed to the onnxruntime Python package.
Also compiles to WebAssembly and has JS bindings (https://github.com/bminixhofer/tractjs).
IMO the most important crate for working with Neural Networks in Rust. Support for inference of other ML models such as Decision Trees is also coming up.