search

lorenmt / reco

The implementation of "Bootstrapping Semantic Segmentation with Regional Contrast" [ICLR 2022].
https://shikun.io/projects/regional-contrast
Other
162 stars 25 forks source link

question about weight of reco loss #14

Closed hxshust closed 2 years ago

hxshust commented 2 years ago

hi,thank you for your impressive work! I have a question about weight of reco loss.In the paper, total_loss = reco_loss + sup_loss, but I found in my experiment that the value of reco_loss is dozens of times that of sup_loss. If reco_loss is set with a small weight so that it and sup_loss reach an order of magnitude, will it be better? Looking forward to your reply.

lorenmt commented 2 years ago

Hi,

Yes this makes sense. A small weighting on the reco loss in this case can be helpful. But still, you may need to do a hyper-paramter search to see the effect of performance with different task weightings.

Let me know if you have further questions.