STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking [NeurIPS 2025]

🚀 Quick Start

BrainCog Installation

For the BrainCog framework, we recommend installing it via GitHub. You can use the following command in your terminal to install it from GitHub:

pip install git+https://github.com/braincog-X/Brain-Cog.git

Very First Step for STEP

STEP installation

git clone https://github.com/Fancyssc/STEP.git

Start Spikformer Training on CIFAR10 as the "Hello-world" Demo.

conda activate [your_env]
python train.py --config configs/spikformer/cifar10.yml

📂 DeepDive Guides

For specific tasks, completed models and supported datasets, please refer to the corresponding submodule guides:

Spiking-Transformer-Benchmark/
├── cls/               # Classification submodule
│   ├── README.md      
│   └── ...
├── seg/               # Segmentation submodule 
│   ├── README.md      
│   └── ...
├── det/               # Object detection submodule 
│   ├── README.md      
│   └── ...
└── README.md

🔑 Key Features of STEP

One-stop benchmark for Spiking-Transformer research—classification, segmentation, and detection share the same training & evaluation pipeline.

Plug-and-play modules (neurons, encodings, attention, surrogate gradients, heads) let you prototype new ideas without touching the core loop.
Ready-made loaders cover ImageNet, CIFAR, DVS-CIFAR10, N-Caltech101 …
Task adapters integrate with MMSeg and MMDet, so dense prediction experiments need only a config tweak.
Backend-agnostic code runs on SpikingJelly, BrainCog, or BrainPy, and every config is version-locked for full reproducibility. neuromorphic vision systems.

Supported Datasets

cls

Static(e.g. CIFAR/ImageNet...)/Neuromorphic(e.g. CIFAR10-DVS)/3D Point Cloud(e.g. ModelNet40) classification datasets are supported.

See cls/README.md for details.

seg/det

Frequently used datasets for both tasks which are assembled by MMSeg and MMDet are supported.

See MMSegmentation and MMDetection for details.

Resources

Tutorial

The brief tutorial of STEP can be found here.

Checkpoints

The main experimental results, including the corresponding log files, configuration files, and checkpoints, can be downloaded here.

Integrated Neurons

Neuron Node	abbreviation	Neuron Node	abbreviation
Integrate-and-Fire Neuron	IF	Leaky Integrate-and-Fire Neuron	LIF
Parametric Leaky Integrate-and-Fire Neuron	PLIF	Exponential Integrate-and-Fire Neuron	EIF
Integer LIF Neuron	I-LIF	Normarlized Interger LIF Neuron	NI-LIF
Hybrid Dynamics LIF Neuron	HD-LIF	Gated LIF Neuron	GLIF
k-based LIF Neuron	KLIF	Complementary LIF Neuron	CLIF
Parallel Spiking Neuron	PSN	Hodgkin-Huxley Neuron	HHNode
Izhikevich Neuron	IzhNode

Supported Encoding

When working with static datasets, SNNs typically require encoding of static images. Our framework supports various encoding methods, including direct, rate, TTFS, and phase encoding.

📝Citation

@misc{shen2025stepunifiedspikingtransformer,
      title={STEP: A Unified Spiking Transformer Evaluation Platform for Fair and Reproducible Benchmarking}, 
      author={Sicheng Shen and Dongcheng Zhao and Linghao Feng and Zeyang Yue and Jindong Li and Tenglong Li and Guobin Shen and Yi Zeng},
      year={2025},
      eprint={2505.11151},
      archivePrefix={arXiv},
      primaryClass={cs.NE},
      url={https://arxiv.org/abs/2505.11151}, 
}

📦 Version & Changelog

[1.1.0] – 2025-07-03

Added

Support for 3D cloud point classification.
Some known bugs fixed.

[1.0.0] – 2025-5-18

Added

initial version released.

💡Acknowledgement

Thanks to the BrainCog for providing the core ideas and components for this repository.

A full list of contributors can be found here.