SCommander

Spiking Transformer with SSM-Attention Hybrid for Speech Command Recognition.

Layout

.
├── scommander/        Python package (data, models, training, utils, ...)
├── scripts/           Setup + verification + anonymization
│   ├── setup.sh
│   ├── verify_env.py
│   └── anonymize_repo.sh
├── configs/           Experiment configs (OmegaConf YAML)
│   ├── base.yaml
│   ├── dataset/       Per-dataset overrides (SHD, SSC, GSC)
│   └── variant/       Per-variant overrides
├── tests/             Pytest suite (env, LIF forward, dataset shapes)
├── pyproject.toml
├── requirements.txt
└── README.md

Setup

One-button installer (creates conda env scommander, installs CUDA 12.8 toolkit, torch 2.7 cu128 wheels, spiking stack, mamba-ssm from source for Blackwell sm_120):

bash scripts/setup.sh

Flags:

• --no-ssm — skip mamba-ssm/causal-conv1d/cupy (CPU-friendly dev; Track C falls back to pure-PyTorch SSM)
• --skip-cuda — skip conda CUDA toolkit install (use system nvcc or module load)
• --env-name=NAME — custom conda env name
• --python=3.10 — custom Python version (default 3.11)

Verify post-install:

conda activate scommander
python scripts/verify_env.py
pytest -q

Target hardware

Production training: RTX 5090 32GB (Blackwell sm_120, driver ≥570, CUDA 12.8). GPU may be shared via NVIDIA MPS — effective VRAM ≈ 26GB.

Dev: any CUDA-capable GPU with --no-ssm flag for scaffold smoke tests.

Datasets

SHD, SSC, and Google Speech Commands v2 (GSC). Downloaded automatically via tonic on first training run. Cache path: data/tonic_cache/ (gitignored).

License

MIT. Paper pending; double-blind anonymization via scripts/anonymize_repo.sh.