search

cwq159 / PyTorch-Spiking-YOLOv3

A PyTorch implementation of Spiking-YOLOv3. Two branches are provided, based on two common PyTorch implementation of YOLOv3(ultralytics/yolov3 & eriklindernoren/PyTorch-YOLOv3), with support for Spiking-YOLOv3-Tiny at present.
GNU General Public License v3.0
209 stars 58 forks source link
ann-to-snn channel-wise-normalization coco convert-operators eriklindernoren parameter-normalization pascal-voc pytorch-yolov3 snn spiking-neural-network spiking-yolo ultralytics yolov3-tiny

readme

PyTorch-Spiking-YOLOv3

A PyTorch implementation of Spiking-YOLOv3, based on the PyTorch implementation of YOLOv3(ultralytics/yolov3), with support for Spiking-YOLOv3-Tiny at present. The whole Spiking-YOLOv3 will be supported soon.

Introduction

For spiking implementation, some operators in YOLOv3-Tiny have been converted equivalently. Please refer to yolov3-tiny-ours(*).cfg in /cfg for details.

Conversion of some operators

Usage

Please refer to ultralytics/yolov3 for the basic usage for training, evaluation and inference. The main advantage of PyTorch-Spiking-YOLOv3 is the transformation from ANN to SNN.

Train

$ python3 train.py --batch-size 32 --cfg cfg/yolov3-tiny-ours.cfg --data data/coco.data --weights ''

Test

$ python3 test.py --cfg cfg/yolov3-tiny-ours.cfg --data data/coco.data --weights weights/best.pt --batch-size 32 --img-size 640

Detect

$ python3 detect.py --cfg cfg/yolov3-tiny-ours.cfg --weights weights/best.pt --img-size 640

Transform

$ python3 ann_to_snn.py --cfg cfg/yolov3-tiny-ours.cfg --data data/coco.data --weights weights/best.pt --timesteps 128

For higher accuracy(mAP), you can try to adjust some hyperparameters.

Trick: the larger timesteps, the higher accuracy.

Results

Here we show the results(mAP) of PASCAL VOC & COCO which are commonly used in object detection，and two custom datasets UAV & UAVCUT. dataset yolov3 yolov3-tiny yolov3-tiny-ours yolov3-tiny-ours-snn
UAVCUT 98.90% 99.10% 98.80% 98.60%
UAV 99.50% 99.40% 99.10% 98.20%
VOC07+12 77.00% 52.30% 55.50% 55.56%
COCO2014 56.50% 33.30% 38.70% 29.50%

From the results, we can conclude that: 1) for simple custom datasets like UAV & UAVCUT, the accuracy of converting some operators is nearly equivalent to the original YOLOv3-Tiny; 2) for complex common datasets like PASCAL VOC & COCO, the accuracy of converting some operators is even better than the original YOLOv3-Tiny; 3) for most datasets, our method of transformation from ANN to SNN can be nearly lossless; 4) for rather complex dataset like COCO, our method of transformation from ANN to SNN causes a certain loss of accuracy(which will been improved later).

UAVCUT

avatar

UAV

avatar

PASCAL VOC

avatar

COCO

avatar

References

Articles