search

alinlab / ifseg

IFSeg: Image-free Semantic Segmentation via Vision-Language Model (CVPR 2023)
78 stars 9 forks source link

Unable to reproduce results #5

Closed zyuanbing closed 9 months ago

zyuanbing commented 12 months ago

Excellent work! Thanks for the release.

I followed the instruction here to perform image-free training on ADE20k val, but got extremely low mIoU results(0.0001). Below is my training log:

logging.log

Is there any bug or possible wrong step in the training procedure that I performed? Or any hint?

Thanks in advance.

kami93 commented 10 months ago

Hi. We have checked the codebase, however, the ADE20k results can be correctly reproduced as suggested in the following logs. logging.log

Perhaps the wrong checkpoint file is being loaded in your case. When we run the ADE experiment with the Chinese-version OFA-base checkpoint, we could reproduce a similar log as given in your question. Please check.

Thanks!