PyTorch re-implementation of YOLOv4 architecture (CSPDarknet53+SPP+PANet) without prediction heads.
Usage:
from pytorch_headless_YOLOv4 import _load
model = _load('$path to state dict file$')Pretrained weights provided by AlexeyAB/darknet, converted to .pth format: https://drive.google.com/open?id=1je_EaNMgkLK-h1P_8XspwDjwGOYdaomT
CUDA-optimized implementation of Mish (provided by thomasbrandon/mish-cuda) is used if installed.
Source code: https://github.com/AlexeyAB/darknet
References:
- https://github.com/Tianxiaomo/pytorch-YOLOv4
- https://github.com/romulus0914/YOLOv4-PyTorch
- https://github.com/thomasbrandon/mish-cuda
@article{yolov4,
title={YOLOv4: YOLOv4: Optimal Speed and Accuracy of Object Detection},
author={Alexey Bochkovskiy, Chien-Yao Wang, Hong-Yuan Mark Liao},
journal = {arXiv},
year={2020}
}