Tracel SDK Example (MNIST)

This repository shows how to adapt the standard Burn MNIST example into a Tracel SDK project. It is a single executable that trains the model and reports the experiment (metrics, checkpoints, and artifacts) to Tracel.

What This Example Covers

• defining a custom model artifact with BundleEncode and BundleDecode
• exposing training configuration through MnistTrainingConfig
• creating and running an experiment with Context::cloud() and ExperimentRun
• wiring metrics, checkpoints, and interruption handling through ExperimentRun

Project Layout

• src/main.rs: entry point that opens a cloud Context, creates the MNIST_Training experiment, and runs it
• src/training.rs: training loop, evaluation, and artifact upload
• src/model.rs: model definition and artifact bundle serialization
• src/data.rs: MNIST batching and data augmentation

Run

Context::cloud() needs Tracel credentials. Either set TRACEL_API_KEY in your environment, or authenticate once with:

tracel login

The project's namespace and name come from tracel.toml. By default the example uses the FlexDevice backend; optional backends can be enabled with Cargo features.

cargo run
cargo run --features wgpu
cargo run --features cuda

The run is reported as the MNIST_Training experiment.

More Details

• Tracel SDK docs: docs.rs/tracel