rotatable yolov3

Introduction

A rotatable yolov3 model which can regress the angle of the bounding box

Features

• Advanced neural network models
• Flexible and efficient toolkit(See woodsgao/pytorch_modules)
• Online data augmenting(By imgaug)
• Mixed precision training(If you have already installed apex)
• Efficient distributed training(0.8x faster when using two 2080ti)
• Add a script to convert to caffe model(By woodsgao/pytorch2caffe)

Installation

git clone https://github.com/woodsgao/rotatable_yolov3
cd rotatable_yolov3
pip install -r requirements.txt

Tutorials

Create custom data

Please organize your data in coco format(by default):

data/
    <custom>/
        images/
        coco.json
        train.json
        val.json

You can use split_coco_json.py from woodsgao/cv_utils to split your coco.json file into train.json and val.json

Training

python3 train.py data/<custom>

Distributed Training

python3 -m torch.distributed.launch --nproc_per_node=<nproc> train.py data/<custom>

Testing

python3 test.py data/<custom>/val.json --weights weights.pth

Inference

python3 inference.py data/samples outputs --weights weights.pth

Export to caffe model

python3 export2caffe.py weights/best.pt -nc 21 -s 416 416