interpreter spine

Interpreter spine (Phase 1)

Phase 1 lifts the models onto the Neuromorphic Intermediate Representation (NIR) and proves the translation. A model is declared once as a TopologySpec and rendered twice: into the snnTorch module that trains, and into the nir.NIRGraph that exports.

Topology presets

Pick a topology by name — the training server takes a topology field and the CLI takes --topology:

Preset	Shape
fc_legacy	The original two-layer FC LIF SpikingNet (default)
fc_small	Small FC LIF with an explicit flatten entry stage
conv_net	Conv/pool feature extractor with a linear LIF readout
recurrent_net	FC LIF with a one-step delayed feedback edge

fc_legacy stays the default and keeps SpikingNet's _fc1/_lif1/_fc2/_lif2 state-dict keys, so existing checkpoints load and infer unchanged.

Neuron registry

spikeforge/neurons/ maps a neuron name to its snnTorch factory and its canonical NIR parameter contract. The registry ships leaky, lapicque, synaptic, and recurrent (RLeaky) neurons.

NIR export, interpretation, and validation

• to_nir(spec, module) exports a JSON-able nir.NIRGraph; graph_summary describes its nodes and edges.
• NirInterpreter executes the exported graph independently of snnTorch, so validate(spec, module, spikes) reports a genuine ValidationReport (per-layer max/mean/relative error plus spike agreement, and an overall within_tolerance flag). Every shipped preset validates with bit-exact spikes.
• All nir/nirtorch imports are confined to nir_bridge/api.py.

Verify CLI

Headless export and validate commands. validate exits non-zero when a report falls outside tolerance, so it doubles as a CI gate:

python -m spikeforge.cli.verify export --topology conv_net
python -m spikeforge.cli.verify export --topology fc_legacy --out g.json
python -m spikeforge.cli.verify validate --topology conv_net
python -m spikeforge.cli.verify validate --topology recurrent_net

WebSocket actions

The server answers two new client actions: nir_export returns the graph summary for the active or configured topology, and nir_validate returns a drift report for the active model.