search

rstrudel / segmenter

[ICCV2021] Official PyTorch implementation of Segmenter: Transformer for Semantic Segmentation
MIT License
846 stars 176 forks source link

How to train with multi gpu #13

Closed BruceYu-Bit closed 3 years ago

BruceYu-Bit commented 3 years ago

I wonder how to train with multi-gpu

chenyangh commented 3 years ago

On a single machine, I simply used model = torch.nn.DataParallel(model), and it works. But this doesn't scale very well.

BruceYu-Bit commented 3 years ago

Thanks! Now i met a new question, i wonder a pretain model with smaller patch size,like 4 or 8. Can u provide?

------------------ 原始邮件 ------------------ 发件人: "rstrudel/segmenter" @.>; 发送时间: 2021年7月28日(星期三) 下午2:52 @.>; @.**@.>; 主题: Re: [rstrudel/segmenter] How to train with multi gpu (#13)

On a single machine, I simply used model = torch.nn.DataParallel(model), and it works. But this doesn't scale very well.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

rstrudel commented 3 years ago

About 8x8 patches please see #3 . The multi GPU training should work out of the box with slurm but maybe you have another usecase which does not work as is. I am closing this issue but feel free to reopen if you still meet issues with training.