What it does Hebel trains feed-forward neural networks for classification and regression, using CUDA acceleration via PyCUDA. It bundles the standard 2014 toolkit: dropout, L1/L2 weight decay, SGD with regular and Nesterov momentum, plus YAML-driven configuration files for running experiments.

The interesting bit The author, Hannes Bretschneider, formally disowned it in the README itself — “I no longer actively develop Hebel” — and pointed readers to Chainer. That makes Hebel a small time capsule: a pre-TensorFlow, pre-PyTorch attempt at GPU deep learning in Python, back when “GPU acceleration” still meant writing PyCUDA kernels by hand.

Key highlights

• Feed-forward nets only; CNNs, autoencoders, and RBMs were “planned for the future” and apparently never arrived
• Configuration-driven training via YAML files in examples/
• pip-installable (pip install hebel) with docs on ReadTheDocs
• Published with a Zenodo DOI for academic citation
• Runs on Linux and Windows; macOS “probably” works but was never tested

Caveats

• The author explicitly recommends against using it
• No commits or active maintenance; dependencies like PyCUDA and skdata are increasingly dated
• “Planned” models (CNNs, RBMs) never materialized per the README

Verdict Curious historians of deep-learning tooling might clone it to see how pre-framework GPU nets felt. Anyone who actually needs to train a neural network in 2024 should follow Bretschneider’s own advice and look elsewhere.