tensorflow2.0-yolov3
tensorflow2.0-yolov3 copied to clipboard
tensorflow2.0 implementation of Yolov3
tf2-yolo3
Introduction
A Tensorflow2.0 implementation of YOLOv3
Quick Start
- Download yolov3.weights and darknet53.conv.74 from YOLO website.
- Download COCO dataset
- Modify the dataset root and weights root in the config file
python main_coco.py --resume load_yolov3 --do_test --net_size 608 --batch_size 8
Training
- run the following command to start training
python main_voc.py/main_coco.py --resume load_darknet --net_size 480 --batch_size 8
- The ckpt file will be stored in the ./checkpoints/dummy_name named with epoch number and use --resume xyz to resume from the xyz epoch
Visualization
The Tensorboard is origanized like TF-ObjectDection-API
Performance
Model | Initial weight | basic resolution | VOC2007 Test(mAP) |
---|---|---|---|
YoloV3 | Darknet53 | 512 | 0.7796 |
Supported Attributes
- [x] Data agumentation
- [x] Multi-scale Training
- [x] Multi-scale Testing(including flip)
- [x] Focal loss
- [ ] ....
TODO
- [x] Update VOC performance
- [ ] Update COCO performance
- [ ] Support distribute training
- [ ] Support Custom dataset