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

The performance obtained is not ideal #15

Open zxzxzxZZZ opened 1 year ago

zxzxzxZZZ commented 1 year ago

Hello, thank you for your excellent work. Can you provide a detailed environment configuration for running your code? The results I achieved locally differ significantly from the expectations you provided. image

GiscardBiamby commented 1 year ago

Did you try assigning any of the models or inputs to a particular CUDA device, or did you set CUDA_VISIBLE_DEVICES environment var? I don't know why but when I did that I got results like yours. When I change the demo notebook back to original and don't change any CUDA device settings I get results that match the expected result.

zxzxzxZZZ commented 1 year ago

Thank you for your reply, but I haven't made any modifications.

yan-xu-helm commented 1 year ago

Are you seeing message like CUDA OOM --> CPU? I think when you run on a machine without enough vRAM, the model will be moved to CPU and weird things happen. I would suggest running on A100 instance and see if you can get the expected results