Home

Awesome

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.

datasetyolov3yolov3-tinyyolov3-tiny-oursyolov3-tiny-ours-snn
UAVCUT98.90%99.10%98.80%98.60%
UAV99.50%99.40%99.10%98.20%
VOC07+1277.00%52.30%55.50%55.56%
COCO201456.50%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

GitHub