Burn Train

Training and evaluation for Burn models

Applications enable the train feature of burn and use this crate as burn::train:

burn = { version = "0.22", features = ["train", "wgpu"] }

• Learner bundles a model, optimizer and learning rate scheduler, and SupervisedTraining runs it over training and validation data loaders on one or several devices.
• metric: loss, accuracy, precision and recall, F-scores, AUROC, BLEU, CER, WER, perplexity, system usage and more; with the vision feature, image metrics such as PSNR, SSIM, LPIPS and FID.
• renderer: a terminal dashboard (tui feature) or plain CLI output for training progress.
• checkpoint: periodic and metric-based checkpointing, and resuming from a checkpoint.
• Early stopping, interruption and an Evaluator for test sets.

Failures on a device during training are returned as errors rather than panics.

See the learner and metric chapters of the Burn Book, and the guide for a complete example.

Feature Flags

• tui (default): terminal dashboard.
• sys-metrics (default): CPU, memory and GPU usage metrics.
• vision: image quality metrics.
• rl: reinforcement learning training through burn-rl.
• tracing: instrument operations with the tracing crate.

Part of the Burn deep learning framework. See the Burn Book and the API documentation.