Rotate - Yolov5

Requirements

• CUDA >= 11.2

Install require package

pip install -r requirements.txt

Install extensions for CUDA

cd utils/iou_cuda
python setup.py install

Dataset

• Use dataset similar to yolov5, but change in label: [cx, cy, w, h, real, imagine]
• Check Convert_labels.ipynb to convert labels and prepare dataset tree.
• Edit path and number of classes in data/rotate_cepdof.yaml

Train

Config

• Edit number of classes in models/rotate_yolov5s_cepdof.yaml
• Run utils.rotate_kmean_anchors to generate predefined anchors for your specific datasets with specified image size.

Training from scratch

python rotate_train.py --weights '' --data data/rotate_cepdof.yaml --cfg models/rotate_yolov5s_cepdof.yaml --rotate

Detect

python rotate_detect.py --weights rotate_best.pt --img 1024 \
--conf 0.75 --source Lunch2_000001.jpg --iou-thres 0.4 

Export ONNX

export ONNX_EXPORT=1
python models/export.py --weights weights/rotate_best.pt --include onnx --dynamic --opset-version 11