jl
jl
> Hi, 我是把连续的几帧的feature拼起来做的。 就像滑动窗口那样做的吗? 还有, 是否考虑公开代码? 我很想知道FAN_use()这个nonID具体是怎么训练的? 论文中介绍的也很简略
公开训练代码
> > 我在运行python web_demo.py --num_gpus 4的时候，执行到第49行代码的时候就报错cuda out of memory退出程序，执行不了后续代码。nvidia-smi监测显示只在第一张显卡上有效，其他显卡都没用上。 > > > > 这种情况怎么处理呢 > > 请问您解决了吗？ > > 我尝试将`device_map='cuda'`改成`device_map='auto'`，模型可以成功加载到多张GPU上，但是会报错`RuntimeError: Expected all tensors to be on the same device, but...
Run into the same problem. It seems all of a suddent both my finetuned model and the official model change to morons. Especially, this is the result by the official...
drives me crazy
As far as I have worked out, It seems that the hf-ds training code is problematic. I guess that is the reason why ds.sh is called "simple ft"... I guess...
> > As far as I have worked out, It seems that the hf-ds training code is problematic. I guess that is the reason why ds.sh is called "simple ft"......