SNNTorch

A second end-to-end example, exported from snnTorch. Where the Oxford example is a real trained model, this one is a decoy network built to exercise the pipeline: it deliberately mixes different node types and is driven by random inputs, so it's useful for testing the NIR→MLIR path and the toolchain rather than for measuring task accuracy.

Network structure

Linear(784 → 256) → LIF(256) → Affine(256 → 10) → CubaLIF(10)

The mix is the point. In a single graph it exercises:

• Linear (fc1, no bias) and Affine (fc2, with bias) — both map to snn.linear, covering the bias / no-bias paths;
• a single-state LIF neuron (lif1) and a two-state CubaLIF neuron (lif2) — exercising both neuron families and, in quantized mode, two snn.rescale insertions.

The 784-wide input matches a flattened 28×28 frame; the random input.h simply drives spikes through this topology.

Files in the example

File	Role
network.nir	The snnTorch-exported network in NIR.
input.h	Pre-baked (random) input (L0_input[N_STEPS][INPUT_SIZE]).
target.csv	Reference output for comparison.
run.py	Driver: exports network.mlir and generates the C runtime into build/.

Run it

uv run python examples/snntorch/run.py              # float32
uv run python examples/snntorch/run.py --quantize   # int8 weights, Q12 state
uv run python examples/snntorch/run.py --n-steps 50 # default is 25

run.py options:

Flag	Default	Effect
--quantize	off	int8 weights + Q12 fixed-point neuron state
--n-steps N	25	Number of simulation timesteps baked into main.c

The generated build/ directory and the lower→compile→run steps are identical to the Oxford example — just swap the paths from examples/snn_oxford/ to examples/snntorch/.