«YOLOv5» implements a tiny version of the original ultralytics/yolov5
ultralytics/yolov5 provides a perfect object detection implementation, including the advanced yolov5 model and loss function, as well as perfect logging and debugging functions. For beginners, training, testing, and deployment of object detection tasks can be completed through documentation, and even ultralytics/yolov5 provides SOP for classification and segmentation tasks.
This repository aims to implement a simplified version of YOLOv5, simplifying the internal implementation of the original YOLOv5 repository as much as possible, and removing features and code that I currently do not need. For example, I will remove the implementation code for video files and cache files in the data module, and only retain the onnxruntime/opencv implementation in the deployment module, and so on.
Note1: The implementation of yolov5 for this warehouse is referenced from v7.0 - YOLOv5 SOTA Realtime Instance Segmentation.
Note2: The configuration of this warehouse is completely based on the original implementation of YOLOv5, divided into configs/data/*.yaml, configs/hyps/*.yaml, configs/models/*.yaml.
- Detect
python -m torch.distributed.run --nproc_per_node 4 --master_port 53122 train.py --data coco.yaml --weights "" --cfg yolov5s.yaml --img 640 --device 0,1,2,3- Classify
python -m torch.distributed.run --nproc_per_node 4 --master_port 25123 classify/train.py --model yolov5s-cls.pt --data imagenet --img 224 --epochs 90 --device 0,1,2,3- Detect
python val.py --weights yolov5s.pt --data coco.yaml --img 640
python val.py --weights yolov5s_voc.pt --data VOC.yaml --img 640| Model | size (pixels) |
dataset |
mAPval 50-95 |
mAPval 50 |
Speed PyTorch RTX3090 (ms) |
params (M) |
FLOPs @640 (B) |
|---|---|---|---|---|---|---|---|
| YOLOv5x | 640 | COCO | 48.1 | 66.7 | 19.7 | 86.71 | 205.5 |
| YOLOv5l | 640 | COCO | 46.2 | 65.2 | 11.7 | 46.53 | 109.0 |
| YOLOv5m | 640 | COCO | 42.9 | 62.1 | 5.0 | 21.17 | 48.9 |
| YOLOv5s | 640 | COCO | 34.7 | 53.8 | 3.6 | 7.23 | 16.4 |
| YOLOv5n | 640 | COCO | 24.4 | 41.3 | 3.5 | 1.87 | 4.5 |
| YOLOv3 | 640 | COCO | 43.6 | 63.7 | 8.0 | 61.92 | 155.9 |
| YOLOv5s | 640 | VOC | 46.8 | 73.8 | 2.3 | 7.06 | 15.9 |
| YOLOv3 | 640 | VOC | 56.9 | 81.9 | 7.1 | 61.60 | 154.9 |
| YOLOv3-Tiny | 640 | VOC | 25.3 | 54.2 | 1.9 | 8.71 | 13.0 |
- Classify
python classify/val.py --weights yolov5s-cls.pt --data ../datasets/imagenet --img 224| Model | size (pixels) |
acc top1 |
acc top5 |
Training 90 epochs 4xRTX3090 (hours) |
Speed PyTorch RTX3090 (ms) |
params (M) |
FLOPs @224 (B) |
|---|---|---|---|---|---|---|---|
| YOLOv5s-cls | 224 | 64.9 | 86.0 | 38.831 | 0.3 | 6.45 | 11.4 |
| YOLOv3-cls | 224 | 68.3 | 88.2 | 56.517 | 1.0 | 16.81 | 98.6 |
| ResNet50 | 224 | 69.3 | 88.2 | 94.422 | 0.4 | 25.6 | 8.5 |
| EfficientNet_b0 | 224 | 71.0 | 90.2 | 77.515 | 0.5 | 5.3 | 1.0 |
python detect.py --weights yolov5n.pt --source assets/coco/
python classify/predict.py --weights yolov5s-cls.pt --source assets/imagenet-val/ --imgsz 224
- zhujian - Initial work - zjykzj
Anyone's participation is welcome! Open an issue or submit PRs.
Small note:
- Git submission specifications should be complied with Conventional Commits
- If versioned, please conform to the Semantic Versioning 2.0.0 specification
- If editing the README, please conform to the standard-readme specification.
Apache License 2.0 © 2023 zjykzj