Bunsen
by ZSpaceLabs
bunsen aims to be a "batteries included" complementary community standard library for extending
the burn tensor library: reusable modules and tensor operations, runtime shape contracts,
pretrained-model loading, and whole models built from them.
- The bunsen book explains how bunsen is organized, the systems that cut across it, and how it is developed. Start with Installing and first use.
- The API docs document every type and lifecycle, with worked examples, and list the crate's features.
Kits
A kit is a whole model or simulation, with its configs and, for a model, the weights it loads by name (Kits):
images::resnet: ResNet image classifiers, with torchvision's andtimm's checkpoints by name.images::swin::v2: Swin Transformer V2 image classifiers, without pretrained weights.gpts::nanochat: a port of karpathy's nanochat GPT, and the corpus it trains on. Work in progress.sims::conway: Conway's Game of Life on toroidal 2D and 3D boards.sims::lbm: a D2Q9 lattice-Boltzmann fluid.speech::whisper: OpenAI's Whisper, with a pretrained index, decoders and a stream driver.speech::silero_vad: the Silero voice-activity detector, with weights a build can bundle.tokens: not a model; the ids-to-text seam the model kits share.
The workspace
Most users depend on bunsen alone. The bunsen repository also publishes its
companion crates (the firehose data pipeline, a Parquet-to-tokens data loader, the shape-contract macro, and bundled
model assets), and holds the development, validation and example crates.
The workspace describes each one.
Examples
The repository's examples/ directory holds runnable
programs, one or more per kit or system. The book's
Examples page says what each one shows.
License
bunsen is distributed under the terms of both the MIT license and the Apache License (Version 2.0).