examples

tract examples

Small end-to-end programs demonstrating how to load and run models with tract. Most use the high-level tract facade; a few still sit on the lower-level framework crates (tract-tensorflow, tract-tflite) and are flagged below.

Image classification

Simple vision pipelines: load a model, feed a pre-processed image tensor, print the top class. Useful entry points when you're just kicking tract's tires.

Example	Model source	Notes
onnx-mobilenet-v2	ONNX	MobileNet v2. Second binary dyn-shape demonstrates symbolic batch dim and runtime shape dispatch.
nnef-mobilenet-v2	NNEF	Same model, pre-compiled to NNEF. Faster load, smaller artifact.
nnef-dump-mobilenet-v2	TF → NNEF	Shows how to tract … dump --nnef a TF model and then run the NNEF.
pytorch-resnet	PyTorch → ONNX	ResNet classification with ndarray broadcasting for per-channel normalisation.
keras-tract-tf2	Keras → ONNX	Numeric equivalence check: runs the same inputs through the ONNX model and compares to .npz reference outputs.

Detection and face similarity

Example	What it shows
face_detection_yolov8onnx_example	YOLOv8-face ONNX model, letterbox pre-processing, bounding-box decoding.
face_similarity_arcface_onnx	Two-stage pipeline: YOLOv8-face for detection, ArcFace for embedding, cosine similarity between two images.

NLP

Example	What it shows
pytorch-albert-v2	ALBERT masked-LM token completion; demonstrates tract's tokenizer integration and tuple-argument run(...).
causal_llm	Reusable crate + CLI + OpenAI-compatible HTTP server for NNEF-packaged causal LLMs (Llama-style). Exercises KV-cache unfolding, prompt chunking, repeat-penalty sampling, state freeze/truncate.

Speech recognition

NVIDIA NeMo RNN-Transducer-style ASR pipelines: preprocessor + encoder + decoder + joint network driven as four separate tract models.

Example	What it shows
nemo-parakeet-asr	Parakeet offline ASR on a WAV file; showcases tract::nnef() loading and CPU/GPU runtime selection.
nemo-nemotron-asr	Nemotron ASR with a transformers_detect_all transform pass and a patch transform to prune a variable-length input.

Image generation (diffusion)

End-to-end text-to-image pipelines with CLIP / T5 text encoders, transformer / UNet backbones, and VAE decoders. All three pick the best available runtime (cuda → metal → cpu).

Example	What it shows
stable-diffusion	SD 1.5 txt2img; Euler scheduler built from scratch, classifier-free guidance, 512×512 output.
stable-diffusion-xl	SDXL 1.0 txt2img; two text encoders, pooled + hidden-state embeddings, time_ids, 1024×1024 output.
stable-diffusion-3	SD 3 Medium txt2img; MMDiT transformer backbone, optional T5-XXL text encoder, flow-matching scheduler.

Ndarray interop

Examples that consume or produce ndarray::ArrayD values call tract::impl_ndarray_interop!() at their crate root. The macro expands in the caller's scope, so the ndarray version used is whichever one the example declares in its Cargo.toml — tract itself no longer re-exports ndarray. See the macro's rustdoc or onnx-mobilenet-v2/src/main.rs for the basic pattern.

Legacy

TensorFlow and TFLite are legacy formats as far as tract is concerned: the tract-tensorflow and tract-tflite crates are still maintained for existing users but the tract facade intentionally doesn't expose them, and that isn't expected to change. Prefer ONNX or NNEF for new work.

Example	Model source	Notes
tensorflow-mobilenet-v2	TensorFlow .pb	Uses tract-tensorflow directly.
tflite-mobilenet-v3	TFLite	Uses tract-tflite directly.