Like my previous project Mobilenet-YOLO , the loss function was very simular to original implemention of darknet
A pytorch implementation of MobileNet-YOLO detection network , train on 07+12 , test on VOC2007 (imagenet pretrained , not coco)
Network | mAP | Resolution | download |
---|---|---|---|
MobileNetV2 | 72.1 | 352 | checkpoint |
- Download dataset VOCdevkit/ , if already have , please skip this step
sh scripts/VOC2007.sh
sh scripts/VOC2012.sh
- Create lmdb
sh scripts/create.sh
- Start training
sh scripts/train.sh
see branch
nnictl create --config config.yml
Download checkpoint, and save at $Mobilenet-YOLO-Pytorch/checkpoint/
sh scripts/inference.sh
- A new detector
- yolov4
- Multi-Task
- Hyper Parameter Tuning
- Pruning
- Porting (KL520 , ncnn , caffe , ...)