---

name: ann2snn description: "ANN-to-SNN conversion workflows for CNNs and Transformers, including FX versus module conversion, rate coding, Transformer TD-equivalent paths, STA, SpikeZIP, Qwen2 calibration, and tiny synthetic validation patterns." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

ANN2SNN

Use this sub-skill when the request is about converting a trained ANN into a SpikingJelly SNN, choosing between FX and module-tree conversion, calibrating with synthetic batches, or validating time-axis readout semantics.

Route here for

• Converter / FXConverter workflows that use RateCodingRecipe, LocalThresholdBalancingRecipe, TransformerTDEquivalentRecipe, or STATransformerRecipe.
• ModuleConverter workflows that use SpikeZIPTFQANNRecipe or Qwen2SNNRecipe.
• Calibration objects and replay contracts such as Qwen2SNNCalibration, SignedQCFSSequenceEncoder, and ChannelVoltageScaler.
• Questions about explicit time-dimension readout, cache continuation, or the difference between dense ANN logits and converted sequence outputs.

Do not use for

• General SNN modeling, neuron/reset semantics, or step-mode basics; route to core-snn.
• Dataset acquisition, file layout, or calibration-data preparation; route to datasets.
• CuPy/Triton/FP8/backend profiling or kernel behavior; route to performance-and-analysis.

What this skill owns

• CNN ANN-to-SNN conversion with rate coding and local threshold balancing.
• Transformer ANN-to-SNN conversion through TD-equivalent FX rewriting.
• STA conversion for Transformers, including calibration-driven spike encoders.
• SpikeZIP QANN-to-SNN module-tree conversion.
• Tiny Qwen2 calibration, conversion, replay, cache, and generation flows.
• Synthetic smoke patterns that validate the converter choice and readout rule without downloading data or checkpoints.

Fast routing rules

1. If the model is a CNN or plain feedforward ANN, start with RateCodingRecipe or LocalThresholdBalancingRecipe on the FX path.
2. If the model is a Transformer and you want FX graph rewriting, use TransformerTDEquivalentRecipe for the lightweight TD baseline or STATransformerRecipe for calibrated STA conversion.
3. If the model is already SpikeZIP-compatible QANN, use ModuleConverter plus SpikeZIPTFQANNRecipe.
4. If the model is Hugging Face Qwen2, calibrate first with calibrate_qwen2_snn, then convert with ModuleConverter and Qwen2SNNRecipe.
5. For any step-mode or reset question, follow core-snn and keep this skill focused on conversion contracts.

Bundled files

• references/conversion-recipes.md
• references/troubleshooting.md
• scripts/ann2snn_tiny_smoke.py

Cross-links

• ../core-snn/ for functional.set_step_mode, functional.reset_net, and state-reset semantics.
• ../datasets/ for calibration-data sourcing and loader layout.
• ../performance-and-analysis/ for optional Triton or FP8 backend behavior.

Evidence used

• Source: spikingjelly/activation_based/ann2snn/*
• Tutorials: docs/source/tutorials/en/ann2snn.rst, docs/source/tutorials/en/ann2snn_transformer.rst
• API docs: docs/source/APIs/spikingjelly.activation_based.ann2snn.rst, docs/source/APIs/spikingjelly.activation_based.ann2snn.examples.rst
• Tests: test/activation_based/test_ann2snn.py, test/activation_based/test_ann2snn_operators.py, test/activation_based/test_ann2snn_qcfs.py, test/activation_based/test_ann2snn_transformer.py, test/activation_based/test_ann2snn_qwen2.py
• Live signatures from the prepared inspection environment for the public converters, recipes, Qwen2 calibration objects, TD operators, and estimate_delay_start.

Read this first

• Use the conversion recipes and signatures in the bundled references rather than the source checkout.
• Readout is explicit: the converted model usually returns a time sequence, and the caller decides when to sum or cache it.
• Qwen2 conversion is a module-tree path; it is not driven through the FX converter.