YOLOv4
This is PyTorch implementation of YOLOv4 which is based on ultralytics/yolov3.
development log
Expand
2020-12-18 - support non-local series self-attention blocks. gc dnl2020-12-16 - support down-sampling blocks in cspnet paper. down-c down-d2020-12-03 - support imitation learning.2020-12-02 - support squeeze and excitation.2020-11-26 - support multi-class multi-anchor joint detection and embedding.2020-11-25 - support joint detection and embedding.2020-11-23 - support teacher-student learning.2020-11-17 - pytorch 1.7 compatibility.2020-11-06 - support inference with initial weights.2020-10-21 - fully supported by darknet.2020-09-18 - design fine-tune methods.2020-08-29 - support deformable kernel.2020-08-25 - pytorch 1.6 compatibility.2020-08-24 - support channel last training/testing.2020-08-16 - design CSPPRN.2020-08-15 - design deeper model. csp-p6-mish2020-08-11 - support HarDNet. hard39-pacsp hard68-pacsp hard85-pacsp2020-08-10 - add DDP training.2020-08-06 - support DCN, DCNv2. yolov4-dcn2020-08-01 - add pytorch hub.2020-07-31 - support ResNet, ResNeXt, CSPResNet, CSPResNeXt. r50-pacsp x50-pacsp cspr50-pacsp cspx50-pacsp2020-07-28 - support SAM. yolov4-pacsp-sam2020-07-24 - update api.2020-07-23 - support CUDA accelerated Mish activation function.2020-07-19 - support and training tiny YOLOv4. yolov4-tiny2020-07-15 - design and training conditional YOLOv4. yolov4-pacsp-conditional2020-07-13 - support MixUp data augmentation.2020-07-03 - design new stem layers.2020-06-16 - support floating16 of GPU inference.2020-06-14 - convert .pt to .weights for darknet fine-tuning.2020-06-13 - update multi-scale training strategy.2020-06-12 - design scaled YOLOv4 follow ultralytics. yolov4-pacsp-s yolov4-pacsp-m yolov4-pacsp-l yolov4-pacsp-x2020-06-07 - design scaling methods for CSP-based models. yolov4-pacsp-25 yolov4-pacsp-752020-06-03 - update COCO2014 to COCO2017.2020-05-30 - update FPN neck to CSPFPN. yolov4-yocsp yolov4-yocsp-mish2020-05-24 - update neck of YOLOv4 to CSPPAN. yolov4-pacsp yolov4-pacsp-mish2020-05-15 - training YOLOv4 with Mish activation function. yolov4-yospp-mish yolov4-paspp-mish2020-05-08 - design and training YOLOv4 with FPN neck. yolov4-yospp2020-05-01 - training YOLOv4 with Leaky activation function using PyTorch. yolov4-paspp
Pretrained Models & Comparison
| Model | Test Size | APval | AP50val | AP75val | APSval | APMval | APLval | cfg | weights |
|---|
| YOLOv4 | 672 | 47.7% | 66.7% | 52.1% | 30.5% | 52.6% | 61.4% | cfg | weights |
| | | | | | | | | |
| YOLOv4pacsp-s | 672 | 36.6% | 55.5% | 39.6% | 21.2% | 41.1% | 47.0% | cfg | weights |
| YOLOv4pacsp | 672 | 47.2% | 66.2% | 51.6% | 30.4% | 52.3% | 60.8% | cfg | weights |
| YOLOv4pacsp-x | 672 | 49.3% | 68.1% | 53.6% | 31.8% | 54.5% | 63.6% | cfg | weights |
| | | | | | | | | |
| YOLOv4pacsp-s-mish | 672 | 38.6% | 57.7% | 41.8% | 22.3% | 43.5% | 49.3% | cfg | weights |
| 640 | 39.9% | 59.1% | 43.1% | 24.4% | 45.2% | 51.4% | | weights |
| YOLOv4pacsp-mish | 672 | 48.1% | 66.9% | 52.3% | 30.8% | 53.4% | 61.7% | cfg | weights |
| 640 | 48.3% | 67.2% | 52.7% | 30.8% | 53.8% | 62.4% | | weights |
| YOLOv4pacsp-x-mish | 672 | 50.0% | 68.5% | 54.4% | 32.9% | 54.9% | 64.0% | cfg | weights |
| 640 | 51.0% | 69.7% | 55.5% | 33.3% | 56.2% | 65.5% | | weights |
| | | | | | | | | |
Requirements
pip install -r requirements.txt
※ For running Mish models, please install https://github.com/thomasbrandon/mish-cuda
Training
python train.py --device 0 --batch-size 16 --img 640 640 --data coco.yaml --cfg cfg/yolov4-pacsp.cfg --weights '' --name yolov4-pacsp
Testing
python test.py --img 640 --conf 0.001 --batch 8 --device 0 --data coco.yaml --cfg cfg/yolov4-pacsp.cfg --weights weights/yolov4-pacsp.pt
Teacher-Student Learning
| Model | Teacher | Test Size | APval | AP50val | AP75val | APSval | APMval | APLval |
|---|
| YOLOv4pacsp-s-mish | - | 672 | 38.6% | 57.7% | 41.8% | 22.3% | 43.5% | 49.3% |
| YOLOv4pacsp-s-mish | YOLOv4pacsp-mish | 672 | 39.3% | 58.4% | 42.5% | 23.4% | 44.5% | 50.7% |
| | | | | | | | |
Citation
@article{bochkovskiy2020yolov4,
title={{YOLOv4}: Optimal Speed and Accuracy of Object Detection},
author={Bochkovskiy, Alexey and Wang, Chien-Yao and Liao, Hong-Yuan Mark},
journal={arXiv preprint arXiv:2004.10934},
year={2020}
}
@inproceedings{wang2020cspnet,
title={{CSPNet}: A New Backbone That Can Enhance Learning Capability of {CNN}},
author={Wang, Chien-Yao and Mark Liao, Hong-Yuan and Wu, Yueh-Hua and Chen, Ping-Yang and Hsieh, Jun-Wei and Yeh, I-Hau},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
pages={390--391},
year={2020}
}
Acknowledgements
- https://github.com/AlexeyAB/darknet
- https://github.com/ultralytics/yolov3
- https://github.com/ultralytics/yolov5
