search

mhamilton723 / FeatUp

Official code for "FeatUp: A Model-Agnostic Frameworkfor Features at Any Resolution" ICLR 2024
MIT License
1.33k stars 76 forks source link

Question about SampledCRFLoss #26

Open coolteemf opened 5 months ago

coolteemf commented 5 months ago

Hi, First of all thank you very much for this great contribution, and releasing both the code and pretrained models ! Even more, the code is easy to use and understand, so again, thank you for your efforts.

I have a question about the SampleCRFLoss, is it related to the Conditional Random Fields you mention in the paper ? I have successfully trained a JBU upsampler, with the parameters in jbu_upsampler.yaml, and noticed that the crf loss is used, with a small weight. Looking at the code, I have the intuition that CRFLoss quantifies how the difference in guidance values at 2 locations correlates to the difference in feature values at the same 2 locations, but I'm not sure this is correct. Could you please explain this part or provide a reference ?

Thank you for your time

rraayyii commented 3 months ago

I also have the same question ---- The CRFLoss was not brought up in the paper.

cheny00 commented 3 months ago

Hi, I encountered a "FileNotFoundError: [Errno 2] No such file or directory: '/pytorch-data\cocostuff\curated\train2017\Coco164kFull_Stuff_Coarse.txt'" error message while running the “train_jbu_upsampler.py” file. How did you resolve it? Where do I download the "Coco164kFull_Stuff_Coarse.txt" file? Or what is the way to add the dataset?