Burn Dataset

Datasets, transformations and data sources for Burn

Applications enable the dataset feature of burn (also enabled by train) and use this crate as burn::data::dataset.

• Dataset is random access by index. InMemDataset holds items in memory, and the sqlite feature adds SqliteDataset for datasets larger than memory.
• transform: mapping, sampling, shuffling, windowing, partial views, selection and composition of datasets, applied lazily.
• source: dataset downloads, including Hugging Face datasets.
• vision, nlp, audio: ready-made datasets such as MNIST, CIFAR, image folders, AG News and Speech Commands, behind the features of the same names.

See the dataset chapter of the Burn Book.

Feature Flags

• sqlite: SQLite-backed datasets, using the Turso engine. sqlite-bundled is a deprecated alias.
• vision, nlp, audio: domain datasets and loaders. builtin-sources enables the downloadable vision and NLP datasets.
• dataframe: datasets backed by a Polars DataFrame.
• network: file downloads.
• fake: generated datasets for tests.
• tracing: instrument operations with the tracing crate.

Try the audio dataset with:

cargo run -p burn-dataset --example speech_commands --features audio

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