LorenzoLamberti94/spiking-patches

Spiking Patches — pretrained Transformer classifiers

Weights for a replication of Spiking Patches: Asynchronous, Sparse, and Efficient Tokens for Event Cameras (Øhrstrøm, Güldenring, Nalpantidis, 2025). Code: DTU-PAS/spiking-patches.

These are the Transformer (ViT-B, 89 M params) classifiers on the two gesture-recognition datasets, one checkpoint per input representation: Spiking Patches (SP), Voxel (V), and Frames (F).

⚠️ Coverage: DvsGesture and SL-Animals-DVS only. The paper's third dataset, GEN1 (object detection), is not included here. Checkpoints are weights-only (state_dict, fp32, no optimizer state) — for evaluation/inference, not resuming training.

Results (test accuracy, %)

File	Dataset	Representation	This checkpoint	Paper
DG-T-SP.ckpt	DvsGesture	Spiking Patches	97.73	98.1
DG-T-V.ckpt	DvsGesture	Voxel	97.73	97.7
DG-T-F.ckpt	DvsGesture	Frames	96.21	97.0
SL-T-SP.ckpt	SL-Animals-DVS	Spiking Patches	89.47	91.7
SL-T-V.ckpt	SL-Animals-DVS	Voxel	85.34	90.2
SL-T-F.ckpt	SL-Animals-DVS	Frames	92.48	88.0

DvsGesture reproduces the paper closely (single run each). SL-Animals-DVS is small and high-variance; the checkpoint shipped here is the best of 5 seeds per config. 5-seed mean ± std (test acc): SP 88.3 ± 1.0, Voxel 82.0 ± 3.6, Frames 90.7 ± 0.8.

Usage

Each file maps to a config name in the upstream repo's train.DEFAULT_CONFIGS (filename == config name):

import torch
from huggingface_hub import hf_hub_download
from sp.configs import Config, Dataset
from sp.loaders import load_model

name = "DG-T-SP"                                   # e.g. DG-T-SP, SL-T-V, ...
ckpt = hf_hub_download("LorenzoLamberti94/spiking-patches", f"{name}.ckpt")

# rebuild the exact config used for training
from train import DEFAULT_CONFIGS
_, config = DEFAULT_CONFIGS[name]
model = load_model(config)
model.load_state_dict(torch.load(ckpt, map_location="cpu")["state_dict"])
model.eval()

Datasets must be preprocessed as described in the upstream README before running evaluation.