// the take
A small, personal collection of PyTorch/NumPy notebooks by David Ha (hardmaru) covering mixture density networks and MNIST trained with several evolution strategies variants (CMA-ES, a genetic algorithm, rank-based OES, PEPG) instead of backprop. Useful for someone curious about ES as a training alternative or wanting a compact MDN reference implementation, not for anyone needing a maintained library.
The mnist_es notebooks implement multiple ES variants side by side on the same task, which is a more concrete way to compare CMA-ES vs. GA vs. PEPG than reading the papers in isolation. The mixture density network notebook is a short, readable reference for MDN loss and sampling in PyTorch that's easy to adapt as a template. Coming from hardmaru, who has actual research background in evolution strategies, the implementations are a credible reference rather than copied boilerplate.
Last commit is from April 2018, written against pre-1.0 PyTorch (Variable-era autograd); expect to have to rewrite parts of the code to run on any current PyTorch install. There's no requirements.txt or environment file, so reproducing the exact versions these notebooks were built against is guesswork. The README is a single sentence — no explanation of the ES hyperparameters, what's being compared, or what results to expect, so you're reading code cold to figure out intent. It's effectively abandoned with no activity or fixes for anything that breaks.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →