DDP Batch Size Configs With 8 GPU

[EricZimmermann]

Hello,

I'm training a Faster RCNN with 8 GPUs and it seems like the standard baselines do this with a total batch size of 16. This would lead to 2 images per GPU which seems a little off, seeing as most of the GPU memory goes unused. Should there be any additional scaling done? I would like to replicate the exact baseline (detection) before moving forward.

Thank you!
Eric