spikezoo 0.2.3.3

A deep learning toolbox for spike-to-image models.

📖 About

⚡Spike-Zoo is the go-to library for state-of-the-art pretrained spike-to-image models designed to reconstruct images from spike streams. Whether you're looking for a simple inference solution or aiming to train your own spike-to-image models, ⚡Spike-Zoo is a modular toolbox that supports both, with key features including:

• Fast inference with pre-trained models.
• Training support for custom-designed spike-to-image models.
• Specialized functions for processing spike data.

🚩 Updates/Changelog

• 25-02-02: Release the Spike-Zoo v0.2 code, which supports more methods, provide more usages like training your method from scratch.
• 24-07-19: Release the Spike-Zoo v0.1 code for base evaluation of SOTA methods.

🍾 Quick Start

1. Installation

For users focused on utilizing pretrained models for spike-to-image conversion, we recommend installing SpikeZoo using one of the following methods:

• Install the last stable version 0.2.3 from PyPI:

pip install spikezoo

• Install the latest developing version 0.2.3 from the source code :

git clone https://github.com/chenkang455/Spike-Zoo
cd Spike-Zoo
python setup.py install

For users interested in training their own spike-to-image model based on our framework, we recommend cloning the repository and modifying the related code directly.

git clone https://github.com/chenkang455/Spike-Zoo
cd Spike-Zoo
python setup.py develop

2. Inference

Reconstructing images from the spike is super easy with Spike-Zoo. Try the following code of the single model:

from spikezoo.pipeline import Pipeline, PipelineConfig
import spikezoo as sz
pipeline = Pipeline(
    cfg=PipelineConfig(save_folder="results",version="v023"),
    model_cfg=sz.METHOD.BASE,
    dataset_cfg=sz.DATASET.BASE 
)

You can also run multiple models at once by changing the pipeline (version parameter corresponds to our released different versions in Releases):

import spikezoo as sz
from spikezoo.pipeline import EnsemblePipeline, EnsemblePipelineConfig
pipeline = EnsemblePipeline(
    cfg=EnsemblePipelineConfig(save_folder="results",version="v023"),
    model_cfg_list=[
        sz.METHOD.BASE,sz.METHOD.TFP,sz.METHOD.TFI,sz.METHOD.SPK2IMGNET,sz.METHOD.WGSE,
        sz.METHOD.SSML,sz.METHOD.BSF,sz.METHOD.STIR,sz.METHOD.SPIKECLIP,sz.METHOD.SSIR],
    dataset_cfg=sz.DATASET.BASE,
)

Having established our pipelines, we provide following functions to enjoy these spike-to-image models.

• I. Obtain the restoration metric and save the recovered image from the given spike:

# 1. spike-to-image from the given dataset
pipeline.infer_from_dataset(idx = 0)

# 2. spike-to-image from the given .dat file
pipeline.infer_from_file(file_path = 'data/scissor.dat',width = 400,height=250)

# 3. spike-to-image from the given spike
import spikezoo as sz
spike = sz.load_vidar_dat("data/scissor.dat",width = 400,height = 250)
pipeline.infer_from_spk(spike)

• II. Save all images from the given dataset.

pipeline.save_imgs_from_dataset()

• III. Calculate the metrics for the specified dataset.

pipeline.cal_metrics()

• IV. Calculate the parameters (params,flops,latency) based on the established pipeline.

pipeline.cal_params()

For detailed usage, welcome check test_single.ipynb and test_ensemble.ipynb.

3. Training

We provide a user-friendly code for training our provided base model (modified from the SpikeCLIP) for the classic REDS dataset introduced in Spk2ImgNet:

from spikezoo.pipeline import TrainPipelineConfig, TrainPipeline
from spikezoo.datasets.reds_base_dataset import REDS_BASEConfig
from spikezoo.models.base_model import BaseModelConfig
pipeline = TrainPipeline(
    cfg=TrainPipelineConfig(save_folder="results", epochs = 10),
    dataset_cfg=REDS_BASEConfig(root_dir = "spikezoo/data/REDS_BASE"),
    model_cfg=BaseModelConfig(),
)
pipeline.train()

We finish the training with one 4090 GPU in 2 minutes, achieving 32.8dB in PSNR and 0.92 in SSIM.

🌟 We encourage users to develop their models with simple modifications to our framework, and the tutorial will be released soon.

We retrain all supported methods except SPIKECLIP on this REDS dataset (training scripts are placed on examples/train_reds_base and evaluation script is placed on test_REDS_base.py), with our reported metrics as follows:

Method	PSNR	SSIM	LPIPS	NIQE	BRISQUE	PIQE	Params (M)	FLOPs (G)	Latency (ms)
tfi	16.503	0.454	0.382	7.289	43.17	49.12	0.00	0.00	3.60
tfp	24.287	0.644	0.274	8.197	48.48	38.38	0.00	0.00	0.03
spikeclip	21.873	0.578	0.333	7.802	42.08	54.01	0.19	23.69	1.27
ssir	26.544	0.718	0.325	4.769	28.45	21.59	0.38	25.92	4.52
ssml	33.697	0.943	0.088	4.669	32.48	37.30	2.38	386.02	244.18
base	36.589	0.965	0.034	4.393	26.16	38.43	0.18	18.04	0.40
stir	37.914	0.973	0.027	4.236	25.10	39.18	5.08	43.31	21.07
wgse	39.036	0.978	0.023	4.231	25.76	44.11	3.81	415.26	73.62
spk2imgnet	39.154	0.978	0.022	4.243	25.20	43.09	3.90	1000.50	123.38
bsf	39.576	0.979	0.019	4.139	24.93	43.03	2.47	705.23	401.50

4. Model Usage

We also provide a direct interface for users interested in taking the spike-to-image model as a part of their work:

import spikezoo as sz
from spikezoo.models.base_model import BaseModel, BaseModelConfig
# input data
spike = sz.load_vidar_dat("data/data.dat", width=400, height=250, out_format="tensor")
spike = spike[None].cuda()
print(f"Input spike shape: {spike.shape}")
# net
net = BaseModel(BaseModelConfig(model_params={"inDim": 41}))
net.build_network(mode = "debug")
# process
recon_img = net(spike)
print(recon_img.shape,recon_img.max(),recon_img.min())

For detailed usage, welcome check test_model.ipynb.

5. Spike Utility

I. Faster spike loading interface

We provide a faster load_vidar_dat function implemented with cpp (by @zeal-ye):

import spikezoo as sz
spike = sz.load_vidar_dat("data/scissor.dat",width = 400,height = 250,version='cpp')

🚀 Results on test_load_dat.py show that the cpp version is more than 10 times faster than the python version.

II. Spike simulation pipeline.

We provide our overall spike simulation pipeline in scripts, try to modify the config in run.sh and run the command to start the simulation process:

bash run.sh

III. Spike-related functions.

For other spike-related functions, welcome check spike_utils.py

📅 TODO

• Support the overall pipeline for spike simulation.
• Provide the tutorials.
• Support more training settings.
• Support more spike-based image reconstruction methods and datasets.

🤗 Supports

Run the following code to find our supported models, datasets and metrics:

import spikezoo as sz
print(sz.METHODS)
print(sz.DATASETS)
print(sz.METRICS)

Supported Models:

Models	Source
tfp,tfi	Spike camera and its coding methods
spk2imgnet	Spk2ImgNet: Learning to Reconstruct Dynamic Scene from Continuous Spike Stream
wgse	Learning Temporal-Ordered Representation for Spike Streams Based on Discrete Wavelet Transforms
ssml	Self-Supervised Mutual Learning for Dynamic Scene Reconstruction of Spiking Camera
ssir	Spike Camera Image Reconstruction Using Deep Spiking Neural Networks
bsf	Boosting Spike Camera Image Reconstruction from a Perspective of Dealing with Spike Fluctuations
stir	Spatio-Temporal Interactive Learning for Efficient Image Reconstruction of Spiking Cameras
base,spikeclip	Rethinking High-speed Image Reconstruction Framework with Spike Camera

Supported Datasets:

Datasets	Source
reds_base	Spk2ImgNet: Learning to Reconstruct Dynamic Scene from Continuous Spike Stream
uhsr	Recognizing Ultra-High-Speed Moving Objects with Bio-Inspired Spike Camera
realworld	recVidarReal2019,momVidarReal2021 in SpikeCV
szdata	SpikeReveal: Unlocking Temporal Sequences from Real Blurry Inputs with Spike Streams

✨‍ Acknowledgment

Our code is built on the open-source projects of SpikeCV, IQA-Pytorch, BasicSR and NeRFStudio.We appreciate the effort of the contributors to these repositories. Thanks for @ruizhao26, @shiyan_chen and @Leozhangjiyuan for their help in building this project.

📑 Citation

If you find our codes helpful to your research, please consider to use the following citation:

@misc{spikezoo,
  title={{Spike-Zoo}: Spike-Zoo: A Toolbox for Spike-to-Image Reconstruction},
  author={Kang Chen and Zhiyuan Ye},
  year={2025},
  howpublished = "[Online]. Available: \url{https://github.com/chenkang455/Spike-Zoo}"
}