HKUST-Aerial-Robotics / SIMPL

SIMPL: A Simple and Efficient Multi-agent Motion Prediction Baseline for Autonomous Driving
MIT License
193 stars 22 forks source link

关于并行计算 #17

Open penglo opened 3 months ago

penglo commented 3 months ago

你好,恭喜您取得如此优秀的成果,我在复现你的代码的时候也是使用了8块3090,但是还是跑不起来您所使用的batchsize ,我研究了一下您使用的多卡训练的代码采用的是数据并行的方式,但是大多数我们用来解决多卡运行显存不足时,通常使用的是模型并行的方式,想请教一下您是否做过这方面的部署呢,如果有,希望您开源一下代码,或者我们进行一些交流,我的邮箱是lipl23@mails.jlu.edu.cn。 Hello,Congratulations on achieving such excellent results. While reproducing your code, I also used 8 RTX 3090 GPUs, but I still couldn't manage to run the batch size you used. After some research, I noticed that your multi-GPU training code uses data parallelism, while most of us typically use model parallelism to solve the issue of insufficient GPU memory. I would like to ask if you have deployed model parallelism in your setup. If so, could you please share your code, or perhaps we could have some discussions on this matter? My email is lipl23@mails.jlu.edu.cn.Best regards.

MasterIzumi commented 3 months ago

@penglo Hi, 本文中的网络比较轻量,应该不会遇到一块卡显存不够存模型而需要模型并行的情况。我们仅提供了DDP数据并行的实现,关于模型并行并没有进行尝试。