lucastabelini/LaneATT

multi-gpu training

Sephirex-X opened this issue · 2 comments

Sephirex-X commented

when using Dataparallel on runner.py got Error
File "Lanatt/lib/models/laneatt.py", line 104, in forward
reg_proposals += self.anchors
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:1 and cuda:0!

lucastabelini commented

I have never tried using DataParallel with LaneATT. Maybe you have to move self.anchors to the current GPU?

Sephirex-X commented

I have never tried using DataParallel with LaneATT. Maybe you have to move self.anchors to the current GPU?

yeah I have managed to move self.anchors to current GPU, however, during the training process only one GPU was activated despite DDP or DP was set correctly.