spikingjelly

Why we included this project

For people building with event-based sensors, spiking neural networks have usually meant awkward, bespoke tooling, and SpikingJelly is an attempt to change that. It wraps SNN building blocks in a PyTorch-native API, so you can define spiking layers, train them with gradient-based methods, and convert pre-trained ANN weights into spiking networks instead of starting over. The project ships with several acceleration backends, plain torch, CuPy, and Triton, which matters when you scale to large models or want real-time throughput on DVS camera streams. Bundled event-based datasets and a beginner-friendly design make it a reasonable on-ramp for newcomers, while hardware deployment and framework exchange support (NIR, Lava) appeal to teams targeting actual neuromorphic chips. For anyone weighing SNNs for low-power or temporal data work, this is a coherent, actively maintained starting point rather than a scattered set of isolated experiments.