search

NVlabs / ODISE

Official PyTorch implementation of ODISE: Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models [CVPR 2023 Highlight]
https://arxiv.org/abs/2303.04803
Other
845 stars 45 forks source link

Performance of the demo varies dramatically with the CUDA device #8

Closed Junyi42 closed 1 year ago

Junyi42 commented 1 year ago

Hi, and thank you for your fantastic work! I have encountered a minor issue that I'd like to bring to your attention. I found that when I change the line model.to(cfg.train.device) to model.to("cuda:1") (or any other device) in the demo.ipynb, there is a significant difference in the generated segmentation map compared to the original (please see the attached image below).

output

The original code runs perfectly fine and produces results consistent with those in the paper. However, when I make this modification, I don't encounter any specific warnings or errors, so I'm uncertain where the issue lies (I suspect that perhaps some modules are not loaded correctly). I'd greatly appreciate your help with this issue. Thank you!

xvjiarui commented 1 year ago

Hi @Junyi42 Sorry for the late reply. How about using CUDA_VISIBLE_DEVICES?

Junyi42 commented 1 year ago

Thanks for the reply, and that solves this issue! btw, I think the issue may lie in the loading of the decoder since I found the extracted features from the backbone unchanged in this case.