This is the YOLOv3 implementation using backbone as Mnasnet.
When I searched various yolov3 implementation, I've never seen the case using Mnasnet as a backbone network of YOLOv3.
So I implemented.
The inspiration for this project comes from yjh0410/pytorch-yolo-v2-v3 Thanks.
The goal of this implementation is to be simple, highly extensible, and easy to integrate into your own projects. This implementation is a work in progress -- new features are currently being implemented.
python train.py --dataset "your_voc_dataset_path"
python eval_voc.py --trained_model "your_trained_model.pth"
You can download it from GoogleDrive.
CPU: I7-6700
GPU: GTX 1080Ti 11GB
RAM: DDR4 16GB
Model name | InputSize | TrainSet | TestSet | mAP | Speed on GPU | Speed on CPU |
---|---|---|---|---|---|---|
MnasYOLOv3 | 416x416 | VOC07+12 | VOC07 | 76.15% | 25fps | 2fps |
Applying multi-scale training