image_demo output difference (CUDA_VISIBLE_DEVICES=1 <-> model.to('cuda:1'))

[PigBroA]

Hi

I tried to inference the segmentation model(mask2former_beit_adapter_large_896_80k_cityscapes_ss) and got some weird point.

(1) Run image_demo.py as CUDA_VISIBLE_DEVICES=1 python image_demo.py \ configs/cityscapes/mask2former_beit_adapter_large_896_80k_cityscapes_ss.py \ released/mask2former_beit_adapter_large_896_80k_cityscapes.pth.tar \ data/sample.png

(2) Run image_demo.py as python image_demo.py \ configs/cityscapes/mask2former_beit_adapter_large_896_80k_cityscapes_ss.py \ released/mask2former_beit_adapter_large_896_80k_cityscapes.pth.tar \ data/sample.png \ --device cuda:1

There are 2 kinds of command, commands mean inference with cuda(index: 1), but outputs are different.

Here is (1) output

Here is (2) output

Please let me know the reason If you know what the problem is.

Thank you

[czczup]

It's really weird, and I don't know why at the moment. Usually, I use the first command to inference a single image.

[InterstellarFang]

Hi

I tried to inference the segmentation model(mask2former_beit_adapter_large_896_80k_cityscapes_ss) and got some weird point.

(1) Run image_demo.py as CUDA_VISIBLE_DEVICES=1 python image_demo.py configs/cityscapes/mask2former_beit_adapter_large_896_80k_cityscapes_ss.py released/mask2former_beit_adapter_large_896_80k_cityscapes.pth.tar data/sample.png

(2) Run image_demo.py as python image_demo.py configs/cityscapes/mask2former_beit_adapter_large_896_80k_cityscapes_ss.py released/mask2former_beit_adapter_large_896_80k_cityscapes.pth.tar data/sample.png --device cuda:1

There are 2 kinds of command, commands mean inference with cuda(index: 1), but outputs are different.

Here is (1) output

Here is (2) output

Please let me know the reason If you know what the problem is.

Thank you

@PigBroA

May I ask if your Python version number is 3.7.11 and the torch version number is 1.9.0？ Is the mmcv version number 1.4.2? Thank you!

[PigBroA]

@InterstellarFang I can't remember that exactly. but, I think I used higher versions what you mention. Thanks (maybe, python was 3.8.15, torch was 1.9.1, mmcv was 1.7.0, but I can't make sure that)