longcw/RoIAlign.pytorch

Hi does your roi align layer support multi gpu?

mks0601 opened this issue · 6 comments

mks0601 commented

Hi,
Does your roi align layer support multi gpu?
I got the weird result when RoI align layer from yours.. For example, if I use 2 gpu, the number of batch of the output tensor is 2 times less.

Can you check?

longcw commented

This may be caused by the wrong box_ind. Your input boxes will be split into two tensors when you are using 2 GPUs while the box_ind in the second tensor will not change automatically.

mks0601 commented

Then how can I make box_ind correct?

longcw commented

You can make the values in box_ind corresponding to each GPU when you generate roi.

paucarre commented

@longcw it didn't work for me

boxes_index = boxes_index.cuda(input.data.get_device())

Still it fails saying tensors are in different GPUs. It only works when using the GPU with index 0. So it seems that for some reason I don't understand somewhere the GPU is hardcoded. I've been digging through the code but I couldn't find it unfortunately...

stomachacheGE commented

@paucarre Hi, did you make it work with multiple GPUs?

frostinassiky commented

@stomachacheGE @paucarre I found this cmd works for me:
roi= roi.to(input.device)
The idea is that both feature and roi should be on the same device.