spikeforge

A spiking-neural-network (SNN) toolkit built on snnTorch and PyTorch. Loads MNIST-style and neuromorphic event datasets, encodes them into rate, latency, delta, and random spikes, and trains, validates, exports, and deploys LIF networks — with a live browser dashboard served over WebSockets.

Pre-1.0. Before trusting any number this produces, read Implications and boundaries.

Quickstart

No clone, no Docker — five lines of Python:

pip install spikeforge

from spikeforge import TrainingEngine

engine = TrainingEngine(dataset="mnist", hidden=32, epochs=1, num_steps=5)
for metrics in engine.train():
    last = metrics
print(last)  # {'loss': ..., 'train_accuracy': ..., 'test_accuracy': ...}

That finishes in seconds — about 3 s on a 12-thread desktop CPU, nearer 14 s on a laptop — and lands in the mid-80s for accuracy. The snippet sets no seed, so the exact figure moves between runs, and the test_accuracy it prints is the engine's fast progress probe rather than the whole test split. For numbers measured on the complete held-out split, with the command that reproduces each one, see Benchmarks. The first run downloads MNIST; later runs are offline.

TrainingEngine is the entry point for new code — it owns the training loop, topology selection, encoding, checkpointing, and evaluation. (SNNTrainer, also exported, is the older MNIST rate-coding helper behind the tutorial demo; reach for TrainingEngine.) spikeforge-verify, spikeforge-benchmark, and the NIR/deployment/energy pieces live in Usage.

On a machine with no GPU, install the CPU torch wheels first so pip does not pull the entire CUDA stack in behind them:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu
pip install spikeforge

That is a ~1.1 GB environment instead of ~5.5 GB, and it is the same flow the Dockerfile and CI use. The order matters: with --extra-index-url pip still prefers the CUDA build from PyPI.

Want the live browser dashboard instead? That's the richer, second path:

git clone https://github.com/Capsize-Games/spikeforge.git
cd spikeforge
docker compose up --build

Open http://localhost:8877 — the dashboard connects to the WebSocket on the same host and port. No separate backend or proxy to run.

See Usage and Quickstart for every path (./install.sh, local dev with Vite, examples/, and the spikeforge-* console scripts).

Why not just snnTorch, Norse, Lava, or SpikingJelly?

spikeforge doesn't replace those — it trains through snnTorch and can deploy to Norse and Lava. The question is what it adds on top: a NIR-described interpreter spine with an independent interpreter that re-executes every exported graph and reports numerical drift against the original model (not just export — snnTorch and Norse both export NIR too; Lava-DL currently only reads it, not writes it), a deployment-capability matrix that reports per-target op support and availability instead of just declaring support, an event-driven energy estimator, and a live training/ introspection dashboard. None of the other libraries ship the latter two as part of the library itself, to our knowledge — file an issue if that's stale.

If you already have a training loop in snnTorch/Norse/Lava/SpikingJelly and don't need interop, deployment reporting, energy estimates, or the dashboard, you may not need spikeforge on top of it — that's a fair "obviously no," and better than an uninformed "obviously yes."

Features

• Encoding — rate, latency, delta, and random spike coders.
• Training — fully-connected and convolutional LIF networks with surrogate-gradient cross-entropy, checkpointing, and opt-in AMP / gradient checkpointing / truncated BPTT / multi-GPU.
• Topologies — fc_legacy, fc_small, conv_net, recurrent_net, plus the sequence presets sequence_mlp and sequence_attn.
• Datasets — MNIST, Fashion-MNIST, KMNIST, QMNIST, USPS, EMNIST, CIFAR-10, and (via the events extra) N-MNIST, DVS128 Gesture, CIFAR10-DVS, and Spiking Speech Commands.
• Interpreter spine — NIR export, an independent NIR interpreter, and numerical drift validation.
• Introspection — educational-mode U[t]/I[t]/S[t] traces, trajectory metrics, and surrogate-derivative curves.
• Deployment — a capability matrix, weight quantization with a drift check that can simulate activation/membrane rounding, energy accounting, and executable reference, norse, and lava_loihi2 backends.
• Model hub — a curated, offline-first catalog carrying this project's own trained reference checkpoints (with the accuracy each scores), plus optional live Hugging Face search.
• Dashboard — a React + TypeScript UI with training, introspection, analysis, targets, energy, and hub panels, and seven guided walkthroughs.

What's implemented vs. experimental vs. spec-only

The features list above spans very different levels of maturity. This is the ten-second version; each row links to the honest detail.

What it scores

Reference configurations on the datasets and topologies that ship, each measured on the complete held-out test split, with the command that reproduces it:

Every figure is the complete held-out test split, on a 12-thread x86-64 CPU with no GPU. The full table — epochs, time steps, hardware, seed, and the one command that reproduces each row — is in Benchmarks.

These are reference configurations with stock hyperparameters and a single seed, not tuned attempts at state of the art; read them as a floor the shipped defaults reach, not as a ceiling. The checkpoints they produce are the trained entries in the model hub (spikeforge-hub list --trained).

Packages

This repository is a single workspace that publishes seven distributions, each versioned independently:

The satellites sit at lower version numbers than core by design, not neglect — each package is versioned independently and moves only when it changes, and compatibility.json is the source of truth for which satellite versions go with which core release. If you're pinning versions by hand, read that file rather than assuming semver alignment across packages.

Documentation

This README covers first contact and positioning; it deliberately doesn't go deeper. documentation/ is the full reference — install paths, the CLI tools, architecture, module layout, and the dev workflow — written for contributors and coding agents alike. Also see COOKBOOK.md for copy-pasteable recipes, examples/ for runnable end-to-end scripts, and plans/ for design documents and the roadmap.

See CONTRIBUTING.md and rules.md before opening a pull request.

Citing

If spikeforge is useful in your research, please cite it — see CITATION.cff (GitHub renders a "Cite this repository" button from it automatically).

License

Released under the BSD 3-Clause License — see LICENSE and AUTHORS.