Does it currently support distributed multi card training ？

xinge008 / Cylinder3D

Rank 1st in the leaderboard of SemanticKITTI semantic segmentation (both single-scan and multi-scan) (Nov. 2020) (CVPR2021 Oral)

Apache License 2.0

859 stars 180 forks source link

Does it currently support distributed multi card training ？ #158

Open chenrui17 opened 2 years ago

chenrui17 commented 2 years ago

Will it be supported in the future ？ current single card training cost too much time

L-Reichardt commented 1 year ago

I got the model to run on multiple GPUs, however the training script in this repo is for single GPU.

With current versions of torch / spconv / CUDA the model is a lot faster to train. I rewrote it here for that purpose (for single GPU).

nakatomo8899 commented 1 year ago

How do I run models on multiple GPUs?

L-Reichardt commented 1 year ago

@nakatomo8899 I wrote my own Distributed Data Parallel (DDP) pipeline for this (not open source). I used a combination of Lei Maos cookbook, PyTorch's tutorial, and well documented repos such as Swin in order to do this.