Implementation of the Gated YOLO architecture introduces a dynamic gating mechanism within the YOLO framework to improve efficiency and performance, particularly on edge devices. This novel adaptation focuses on real-time object detection in dynamic environments, leveraging scene-specific model pruning and dynamic gating to significantly reduce computational load without impacting detection accuracy.
Install
git clone https://github.com/hectorandac/Gated-YOLOv6
cd Gated-YOLOv6
pip install -r requirements.txtFinetune on Custom Data
Single GPU
python tools/train.py --batch 32 --conf configs/gated_yolov6s_finetune.py --data data/dataset.yaml --device 0Multi GPUs (DDP mode recommended)
python -m torch.distributed.launch --nproc_per_node 8 tools/train.py --batch 256 --conf configs/gated_yolov6s_finetune.py --data data/dataset.yaml --device 0,1,2,3,4,5,6,7- conf: select config file to specify network/optimizer/hyperparameters. It's recommended to apply gated_yolov6n/s/m/l_finetune.py when training on your custom dataset.
- data: prepare dataset and specify dataset paths in data.yaml
YOLOv6 supports different input resolution modes. For details, see How to Set the Input Size.
Resume Training
If your training process is interrupted, you can resume training by specifying the path to your latest checkpoint:
python tools/train.py --resume /path/to/your/checkpointThis command will resume from the specified checkpoint.
Evaluation
Evaluate model on COCO val2017 dataset with 640×640 resolution
python tools/eval.py --data data/coco.yaml --batch 32 --weights path/to/your/model.pt --task val- config-file: specify a config file to define all the eval params.
Inference
Download a pretrained Gated YOLO model or use your trained model for inference.
python tools/infer.py --weights path/to/your/model.pt --source img.jpg / imgdir / video.mp4To use a local or web camera for inference:
python tools/infer.py --weights path/to/your/model.pt --webcam --webcam-addr 0webcam-addr can be local camera number id or rtsp address.
Tutorials
For any questions or discussions, feel free to open an issue on GitHub.