meituan-gengyifei
commented
2 years ago Thanks for your attention!
- At present, yolo v6 can support the training of single-classification models. If possible, please provide the corresponding error information to facilitate our positioning.
- Regarding the problem that the gpu is not fully utilized, you can refer to #issue159 Apart from that, the reasons for the slower training speed are multifactorial and cannot be attributed to low GPU utilization. On the one hand, we use the structure of repvgg, which is slow in the training process, but fused in the inference process, so the inference speed of yolov6 is very fast. On the other hand, we use the evaluation methods of simota and pycocotools in the network, which play a great role in improving the accuracy of the network and standardizing the evaluation criteria, but it will consume some training time.
- This bug has fixed, Pull the latest code to confirm that your problem still exists.
- At present, we refer to the industry's unified evaluation standards and use the official open source interface of pycocotools to evaluate the map series indicators. After that, we will add other indicators one after another. Welcome to use!
尊敬的作者，我想向您提出一些关于YOLOv6的bug： 1、目前无法使用单类别数据集进行训练 2、训练时GPU无法充分利用，导致训练速度很慢 3、每训练一个epochs，runs中的train文件夹就会新建一个文件 4、能否在训练时添加查看准确率和召回率的信息情况？ 在此放上我总结的一篇博文：https://blog.csdn.net/qq_58355216/article/details/125525243?spm=1001.2014.3001.5502 希望作者尽早修复，我将非常感谢！