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.
For spiking implementation, some operators in YOLOv3-Tiny have been converted equivalently. Please refer to yolov3-tiny-ours(*).cfg in /cfg for details.
- 'maxpool(stride=2)'->'convolutional(stride=2)'
- 'maxpool(stride=1)'->'none'
- 'upsample'->'transposed_convolutional'
- 'leaky_relu'->'relu'
- 'batch_normalization'->'fuse_conv_and_bn'
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.
$ python3 train.py --batch-size 32 --cfg cfg/yolov3-tiny-ours.cfg --data data/coco.data --weights ''
$ python3 test.py --cfg cfg/yolov3-tiny-ours.cfg --data data/coco.data --weights weights/best.pt --batch-size 32 --img-size 640
$ python3 detect.py --cfg cfg/yolov3-tiny-ours.cfg --weights weights/best.pt --img-size 640
$ 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.
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:
- for simple custom datasets like UAV & UAVCUT, the accuracy of converting some operators is nearly equivalent to the original YOLOv3-Tiny;
- for complex common datasets like PASCAL VOC & COCO, the accuracy of converting some operators is even better than the original YOLOv3-Tiny;
- for most datasets, our method of transformation from ANN to SNN can be nearly lossless;
- 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
UAV
PASCAL VOC
COCO
