---

name: core-snn description: "Teach and debug SpikingJelly activation-based core SNN modeling: step modes, state reset, surrogate gradients, neurons, layers, recurrent wrappers, monitors, and tiny timing/visualization helpers." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Core SNN Modeling

Use this sub-skill when the task is to construct, inspect, or debug the core spikingjelly.activation_based object model for activation-based spiking neural networks.

Route here for

• base: step-mode-aware modules, state registration, memory extraction, functional conversion.
• functional: network-wide step-mode/reset/backend helpers and explicit forward wrappers.
• surrogate: Heaviside-based surrogate-gradient modules and gradient helpers.
• neuron: canonical IF/LIF/PLIF/PSN/FlexSN neurons and the main stateful neuron families.
• layer: step-aware stateless layers, recurrent wrappers, stateful synapse helpers, batch-norm variants, and attention wrappers.
• rnn: stacked spiking RNN cells and bidirectional wrappers.
• monitor: output/input/attribute/gradient monitors and GPU sampling.
• timing_based: GaussianTuning and Tempotron.
• visualizing: spike, voltage, and feature-map plotting helpers.
• configure and logger: package-level import-time config and logging behavior.

Do not handle here

• Neuromorphic datasets, download/build flows, and frame/event preprocessing: route to datasets.
• ANN-to-SNN conversion, calibration, and recipe selection: route to ann2snn.
• Backend performance, kernels, precision, profiling, and memory optimization: route to performance-and-analysis.
• Model zoo, training loops, distributed scale-out, and benchmark-heavy workflows: route to training-and-scaleout.
• NIR/Lava/Lynxi exchange or deployment paths: route to deployment-exchange.

Operating workflow

1. Prefer stable public imports from the top-level namespaces:
   • from spikingjelly.activation_based import base, functional, surrogate, neuron, layer, rnn, monitor
   • from spikingjelly.timing_based.encoding import GaussianTuning
   • from spikingjelly.timing_based.neuron import Tempotron
   • from spikingjelly import visualizing
   • from spikingjelly import configure
   • from spikingjelly.logger import logger
2. Decide whether the network is single-step (step_mode='s') or multi-step (step_mode='m'). Use [N, *] for single-step inputs and [T, N, *] for time-major sequences.
3. Reset stateful modules between batches with functional.reset_net(net). If the model structure changes, invalidate the reset cache first with functional.invalidate_reset_cache(net).
4. For custom neuron dynamics, use SimpleBaseNode when you want the readable charge/fire/reset path, or BaseNode when you want explicit functional transitions and backend-aware kernels.
5. Use monitors and plotting helpers only as side-effect-free inspection tools. Remove hooks explicitly when you are done.
6. Keep dataset, ANN2SNN, backend-performance, training, and deployment questions routed out of this sub-skill.

Bundled references

• references/core-api.md
• references/troubleshooting.md

Bundled script

• scripts/minimal_core_smoke.py: tiny no-download smoke for a mixed IF/LIF network, functional.reset_net, functional.set_step_mode, and monitor/surrogate checks.

Evidence base

Consulted source files:

• spikingjelly/activation_based/base.py
• spikingjelly/activation_based/functional/net_config.py
• spikingjelly/activation_based/functional/forward.py
• spikingjelly/activation_based/functional/layer.py
• spikingjelly/activation_based/functional/neuron.py
• spikingjelly/activation_based/functional/misc.py
• spikingjelly/activation_based/surrogate.py
• spikingjelly/activation_based/neuron/base_node.py
• spikingjelly/activation_based/neuron/integrate_and_fire.py
• spikingjelly/activation_based/neuron/lif.py
• spikingjelly/activation_based/neuron/plif.py
• spikingjelly/activation_based/layer/container.py
• spikingjelly/activation_based/layer/stateless_wrapper.py
• spikingjelly/activation_based/layer/bn.py
• spikingjelly/activation_based/layer/misc.py
• spikingjelly/activation_based/rnn.py
• spikingjelly/activation_based/monitor.py
• spikingjelly/timing_based/encoding.py
• spikingjelly/timing_based/neuron.py
• spikingjelly/visualizing/*.py
• spikingjelly/configure.py
• spikingjelly/logger.py

Consulted docs:

• docs/source/tutorials/en/basic_concept.rst
• docs/source/tutorials/en/neuron.rst
• docs/source/tutorials/en/surrogate.rst
• docs/source/tutorials/en/monitor.rst
• docs/source/tutorials/en/recurrent_connection_and_stateful_synapse.rst
• docs/source/APIs/spikingjelly.activation_based.rst
• docs/source/APIs/spikingjelly.activation_based.base.rst
• docs/source/APIs/spikingjelly.activation_based.functional.rst
• docs/source/APIs/spikingjelly.activation_based.neuron.rst
• docs/source/APIs/spikingjelly.activation_based.neuron.core.rst
• docs/source/APIs/spikingjelly.activation_based.neuron.research.rst
• docs/source/APIs/spikingjelly.activation_based.surrogate.rst
• docs/source/APIs/spikingjelly.activation_based.monitor.rst
• docs/source/APIs/spikingjelly.activation_based.rnn.rst
• docs/source/APIs/spikingjelly.timing_based.rst
• docs/source/APIs/spikingjelly.visualizing.rst
• docs/source/APIs/spikingjelly.configure.rst
• docs/source/APIs/spikingjelly.logger.rst

Consulted tests:

• test/test_configure.py
• test/activation_based/test_functional.py
• test/activation_based/test_functional_layer.py
• test/activation_based/test_functional_neuron.py
• test/activation_based/test_monitor.py
• test/activation_based/test_neuron_variants.py
• test/activation_based/test_rnn.py
• test/test_visualizing.py
• test/timing_based/test_encoding.py
• test/timing_based/test_neuron.py
• test/test_logging_policy.py

Verified live-signature evidence:

• skills/tests/spikingjelly/reports/environment/repo_env_report.json confirms Python 3.11, installed spikingjelly 2.0.0.dev1, and required CPU/CUDA/CuPy/Triton/NIR/transformers imports.
• The same report records that the required CUDA/CuPy/Triton smoke passed on the A100 host.