features

Features

• MNIST loading + snntorch.utils.data_subset reduction
• Rate coding (spikegen.rate) at gain 1 and a lower gain
• Latency coding (spikegen.latency) with tau, threshold, linear, normalize, and clip variants (tutorial 2.3)
• Delta modulation (spikegen.delta) with on/off spikes (tutorial 2.4)
• Random spike generation from scratch via spikegen.rate_conv (tutorial 3)
• Matplotlib exports (MP4s, GIFs, rasters, reconstructions) into build/
• Training: a fully-connected LIF spiking network with a surrogate-gradient cross-entropy loss, streaming live loss and both batch + held-out accuracy
• Datasets: train on MNIST, Fashion-MNIST, KMNIST, QMNIST, USPS, EMNIST digits/letters, or (grayscaled) CIFAR-10, all normalised to 28x28 so one architecture fits all
• Model management: save checkpoints, list/load/delete them, and continue training an already-trained model
• Browser interface: a dark-themed grid dashboard that streams encoded spike data over WebSockets and renders plots live in the client, with a training panel (dataset picker, network config, live charts, model management, predictions)
• Compute selection: a CPU/GPU device dropdown (GPU by default, with automatic CPU fallback) and a live CPU-RAM / VRAM resource monitor
• Interpreter spine (Phase 1): topology presets, a neuron registry, NIR export, an independent NIR interpreter, and numerical drift validation (see below)
• Dual-mode introspection (Phase 2): an educational mode that records per-step U[t]/I[t]/S[t], trajectory metrics, encoding/decoding reports, surrogate-gradient curves, a neuron comparison lab, and a production-mode benchmark harness (see below)
• Unified dashboard (Phase 3): an Educational/Production mode toggle, topology/neuron/surrogate pickers, neuron-state trajectory and NIR graph viewers, a drift-validation panel, trajectory-metrics/encoding/surrogate/ benchmark analysis panels, and seven guided walkthroughs (see below)
• Event datasets (Phase 4): N-MNIST, DVS128 Gesture, CIFAR10-DVS, and Spiking Speech Commands through Tonic, with a modality-aware dataset picker, an event-to-spike bridge, and polarity-aware rasters (see below)
• Targets and interoperability (Phase 5): a deployment-target registry with an honest capability matrix, per-target deployment reports, external NIR import/export, and a round-trip fidelity guarantee (see below)
• Production workflows (Phase 6): a reproducibility manifest and config hash, a searchable checkpoint registry with metadata diffing, opt-in training scale-ups (AMP, gradient checkpointing, truncated BPTT, multi-GPU), a stored benchmark suite with regression gating, opt-in JSON logging and a metrics snapshot, packaged console scripts, and Docker CPU/GPU profiles (see below)
• Model hub (WS-A): a bundled curated catalog (10 verified entries across five frameworks) plus optional live Hugging Face access, an isolated downloader with progress/cancel and checksum verification, and an inspect → compat → promote import funnel, surfaced through spikeforge-hub, six WebSocket actions, and the HubPanel browser (see below)
• Backend execution (WS-B): a substitution executor that applies a target's declared rewrites with a report and drift check, and executable reference, norse, and lava_loihi2 backends behind one compile_run entry point, surfaced through deploy/rewrite/run (see below)
• Sequence primitives (WS-C): per-stage heterogeneous neurons, ten new stage kinds with explicit NIR contracts, and the sequence_mlp/sequence_attn demonstration presets (see below)
• Event runtime and energy (WS-D): a sparse/event-driven runner with a dense-parity check, SOP/MAC/AC counting, and a spikeforge-energy report that maps op counts to a declared per-target cost table (see below)
• Operational maturity (WS-E): opt-in persisted metrics, optional TensorBoard/W&B tracking sinks, determinism tooling, and a generated docs site (see below)
• Interop fold-ins (WS-F): event-dataset training, an ONNX bridge, nirtorch extraction of third-party PyTorch modules, weight-level quantization (whose drift check can now also simulate activation/membrane rounding, see Implications and boundaries §3), non-square sensor geometry, and per-step hidden-layer animation (see below)