PyTorch-Spiking-YOLOv3

A modified repository based on PyTorch-Spiking-YOLOv3(cwq159/PyTorch-Spiking-YOLOv3) and YOLOv3(ultralytics/yolov3), which makes it suitable for VOC-dataset and YOLOv2. There is no extra contribution, and thanks for the two authors above. It is feasible for CNN training and testing and SNN testing. For yolov2 or yolov3 training, it is useful to refer ultralytics/yolov3#1570 and get original code as below:

$ git clone https://github.com/ultralytics/yolov3 -b archive  # archive branch

Dataset

VOC dataset should be the same state as PyTorch-Spiking-YOLOv3, such as: /parent /dataset/VOC /PyTorch-Spiking-YOLOv3 Please get VOC.tar in dataset and upzip them. Use voc_label.py in dataset (from https://pjreddie.com/media/files/voc_label.py) to generate labels. Use rebuild_voc.py in VOC to make a new collection of VOCdatset in folder VOC.

Introduction

For spiking implementation, some operators in YOLOv2-Tiny have been converted equivalently. Please refer to yolov3-tiny-ours(*).cfg in /cfg for details.

Conversion of some operators

• 'maxpool(stride=2)'->'convolutional(stride=2)'
• 'maxpool(stride=1)'->'none'
• 'leaky_relu'->'relu'
• 'batch_normalization'->'fuse_conv_and_bn'

Usage

Please refer to ultralytics/yolov3 for the basic usage for training, evaluation and inference.

Train

$ python3 train.py

Test

$ python3 test.py

Detect

$ python3 detect.py

Transform

$ python3 ann_to_snn.py

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

Trick: the larger timesteps, the higher accuracy.