A minimal PyTorch implementation of Spiking-YOLOv3, based on the minimal PyTorch implementation of YOLOv3(eriklindernoren/PyTorch-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 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 eriklindernoren/PyTorch-YOLOv3 for the basic usage of PyTorch-YOLOv3 for training, evaluation and inference. The main advantage of PyTorch-Spiking-YOLOv3 is the transformation from ANN to SNN.

$ python3 train.py

After training, please rename your checkpoint file and move it to the /weights folder.

$ cd checkpoints
$ mv yolov3-tiny-ours_ckpt_99.pth ../weights/yolov3-tiny-ours_best.pth

$ python3 test.py

$ python3 detect.py

$ python3 ann_to_snn.py

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 COCO2014 which is commonly used in object detection，and two custom datasets UAV/UAVCUT.

From the results, we can conclude that:

1. for simple custom datasets, converting some operators is equivalent to the original YOLOv3-Tiny;
2. for complete dataset like COCO2014, the accuracy of converting some operators is lower than the original YOLOv3-Tiny;
3. regardless of datasets, our method of transformation from ANN to SNN can be nearly lossless.

• Theory and Tools for the Conversion of Analog to Spiking Convolutional Neural Networks
• Spiking-YOLO: Spiking Neural Network for Energy-Efficient Object Detection

• NeuromorphicProcessorProject/snn_toolbox
• hahnyuan/ANN2SNN