Simulation and inputs

The graph describes the circuit. Duration, seed, dataset selection, input realisation, output directory, and recording choices belong to the execution protocol.

Graph-native requestsCopy Anchor Link

import torch
from snnlab.sim.execution import ExecutionSpec, simulate

# Assume graph is a compiled bundle's graph mapping.
result = simulate(ExecutionSpec(
    kind="simulate", executor="graph", graph=graph,
    inputs={"events": torch.zeros(100, 1, 128)},
    seed=17, device="cpu",
))

The input must match the named graph contract. Use the complete quickstart to create a graph first. ExecutionResult exposes named outputs, recordings, parameters, final state, and metrics. Typed simulate needs executor="graph" for graph execution; its legacy route returns routing metadata.

Exact dense replayCopy Anchor Link

uv run snnsim sim \
  --executor graph \
  --bundle circuit.bundle \
  --input-file replay.npz \
  --input-dataset-id my-recording-v1 \
  --input-split test \
  --no-input-shuffle \
  --seed 17 \
  --out-dir artifacts/replay

NPY binds to the sole graph input. NPZ arrays bind by input id. Resolution requires exact input coverage, matching time and batch axes, declared feature dimensions, finite values, binary spikes, and boolean or zero/one masks.

Sparse events and datasetsCopy Anchor Link

A single-input event NPZ stores steps, batches, channels, steps_count, and batch_size. Coordinates are zero-based integer steps, ordered by step, batch, and channel. Duplicate or out-of-bounds coordinates are rejected. For multiple inputs, prefix each field with its input id.

DatasetSnapshotBinding consumes an external immutable NPZ snapshot using one of three standard encoders: rate_poisson, prebinned_spikes, or event_bin. Selection, encoding, source digest, labels, split, and seeds are recorded in provenance; the resolver does not download datasets.

Artifacts and continuationCopy Anchor Link

Graph CLI inference writes a digest-bearing inference-manifest.json. Use validate_inference_artifacts from snnlab.sim.execution before reusing a cache. Training checkpoints and resumable runtime state are different contracts: a checkpoint authenticates parameters and training state; runtime continuation carries dynamic simulation state.

Use uv run snnsim sim --help for the complete option set. See scientific contracts before changing timesteps or delays.

API referenceCopy Anchor Link

See the complete snnlab.sim API reference for signatures, defaults, fields, methods and validation.