Mobu59
Mobu59
@GuanLinHu I am the same, I can train with single gpu, but when I use multi-gpu, I meet the problem above!
> Thanks for the amazing work again! > > After replacing the SORT kalman filter in `ocsort.py` with the JDE kalman filter, I got higher HOTA and faster speed, which...
我的也是这样，训练可以正常进行，包括也看得到效果，但是map是0
yolov3.cfg文件第一个卷积层卷积核的数量是32吧，ghostnet的cfg文件中第一个卷积核数量是16，加上第二个卷积核的16就是32了，这里应该是拿一个ghost module替换yolov3的一个卷积层
Have you fixed the problem?
can you give the yaml of yolov6_tiny and yolov6-nano ,thanks a lot@jizhishutong
> can you give the yaml of yolov6_tiny and yolov6-nano ,thanks a lot@jizhishutong solved
hello, have you solved the problem?
Some data seems to be different from the official one? For example, the parameters and flops of yolox-s, the official ones are 9.0 and 26.8, while your experiment is 8.1...
> Yes, but you need use cyclical learning rates for the extra training and then average these 12 checkpoints. hello ,I use the get_swa_model.py and CosineAnnealingLR（initial learning rate 1e-3,eta_min=1e-4）to train...