filaPro / oneformer3d

[CVPR2024] OneFormer3D: One Transformer for Unified Point Cloud Segmentation
Other
348 stars 32 forks source link

About multi-GPU training learning rate #78

Closed RayYoh closed 1 week ago

RayYoh commented 2 months ago

Hi @filaPro, recently, I have been trying to reproduce the results of Oneformer3D to compare with my own method. Since using just one GPU would be relatively slow, I changed to using multi-GPU training to align the configs with my method.
My question is if I use 3 or 4 GPUs for training, which means the batch size will be 34 or 44, do I need to modify the learning rate in the raw configs (i.e. 310-4 or 410-4).

Best

filaPro commented 2 months ago

We never tried multi-gpu for oneforner3d, so probably yes, multiplying should be fine, but i'm not sure if after it the metrics will be reproduced exactly.