How to use multi-gpu training?

[sumitmishra209]

Hello, thanks for the detailed repo of BoQ and https://github.com/amaralibey/OpenVPRLab/tree/main
I have 24 gb 2 GPUs and batch size of 120 can't fit in one GPU memory. So i was thinking to use both, i was trying by adding "devices=[0, 1], strategy='ddp'," in run.py of open vpr lab but got error.

I trained by keeping batch size of 80 on one GPU and accordingly change warmup step but the BOQ result can't reach near to what you got in paper.

[amaralibey]

Hello @sumitmishra209
Thank you for your interest,

I didn't try multi-gpu training as I don't have that kind of setup, sorry.
However, 24GB of memory should be enough to train a DinoV2-BoQ model. You can use batchez of size 160x4, and resize images to 224x224 during training and 322x322 during test. I'm getting ~93.5 R@1 on MSLS-val. You can use OpenVPRLab to track and manage your training (https://github.com/amaralibey/OpenVPRLab).