YOLOX icon indicating copy to clipboard operation
YOLOX copied to clipboard

使用coco预训练模型和resume接着训练有冲突,不可以都加载,请问,如果先使用coco预训练然后停了接着训,同时加载last epoch ckpt和coco预训练权重冲突,这时,不加载coco预训练权重,还有预训练的效果????

Open starsky68 opened this issue 4 years ago • 7 comments

训练几个epoch后出现torch.multiprocessing.spawn.ProcessExitedException: process 1 terminated with signal SIGABRT

starsky68 avatar Oct 21 '21 01:10 starsky68

coco预训练权重在哪里阿?没有找到

charles-str avatar Oct 21 '21 01:10 charles-str

coco预训练权重在哪里阿?没有找到

就在主页了

starsky68 avatar Oct 21 '21 01:10 starsky68

方便加QQ:1877095454,我训练的tiny版本

charles-str avatar Oct 21 '21 01:10 charles-str

Sorry, I could not understand the exact problem you are facing, could you explain it in detail? Any exception backtrack of process 1 terminated with signal SIGABRT ?

FateScript avatar Oct 29 '21 05:10 FateScript

相同的问题,在一张卡上能跑,问题出现在yolox/core/launch.py 的mp.start_processes,偶尔2张卡也能跑

mcmingchang avatar Dec 15 '21 03:12 mcmingchang

相同的问题,在一张卡上能跑,问题出现在yolox/core/launch.py 的mp.start_processes,偶尔2张卡也能跑

我目前是停了后切换模型训练,效果来看,没啥影响,然后你遇到这个多卡不可以跑的问题,可能是你改动了loss或者bs 线程数啥的都需要仔细调整

starsky68 avatar Dec 15 '21 03:12 starsky68

请问有ImageNet的预训练权重吗

zhangchi621 avatar Aug 23 '22 08:08 zhangchi621