MaskDINO Cannot specify GPU when carrying out inference with mask DINO

When I set cfg.MODEL.DEVICE = 'cuda:1' ,it says illegal memory access. The problem does not exist when i set cfg.MODEL.DEVICE = 'cuda:0', how to solve this?

Mar 26 '23 09:03 keithkctse

I am still trying to figure out this problem. If you wanna use cuda:1 and run on this single GPU, you can add 'CUDA_VISIBLE_DEVICES=1' before your command, for example

CUDA_VISIBLE_DEVICES=1  &&  python train_net.py --resume --num-gpus 1 --config-file \
    configs/coco/instance-segmentation/maskdino_R50_bs16_50ep_3s_dowsample1_2048.yaml

This will only use the cuda:1 device.

Mar 29 '23 03:03 FengLi-ust

Hi, thanks for your reply.

We would like to implement multiple maskdino models on multiple gpus, it would be great if we can do so.

Is it because the .cuda() defaults using cuda:0 in the code? Could it be .to('cuda:1') where the 'cuda:1' reads from the cfg.MODEL.DEVICE? Just my guess

Mar 29 '23 03:03 keithkctse

Maybe you guess if right, but I never used your way of specifying cuda devices. You can implement multiple maskdino models on multiple gpus with my command, which works well for me.

Mar 29 '23 03:03 FengLi-ust

We are implementing the models for inference on our project, so probably cannot just use the command above. I want to see if the codes can be modified that it can read from the cfg.MODEL.DEVICE.

https://github.com/IDEA-Research/MaskDINO/blob/95cf05ccfd1d0496bc92980f29a18536c92b450f/maskdino/modeling/transformer_decoder/maskdino_decoder.py#L268

As I see the codes here use 'cuda' only without specifying the gpus, is there a way to pass the gpu no. as specified by cfg.MODEL.DEVICE into those codes?

Mar 29 '23 04:03 keithkctse