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 the contrastive loss #42

Closed Hugo-cell111 closed 11 months ago

Hugo-cell111 commented 11 months ago

Hi! may you provide the training log of the experiment? I want to see the convergence curve of contrastive loss. Because the implemented contrastive loss by myself borrows largely from your code but drops very slightly, I am afraid I might make some mistakes. Thanks a lot!

lorenmt commented 11 months ago

Hi, sorry that I did not record any training logs. But I would suggest having some generated pseudo images for testing. The contrastive loss simply would not go down to a very low value.

Hugo-cell111 commented 11 months ago

Yes, the contrastive loss wouldn't go down to a very low value. But my loss has gone down very slightly(from about 3.84 to 3.8), and I wonder if you have met such situation. Thanks!

----------Reply to Message---------- On Mon, Oct 2, 2023 00:59 AM Shikun @.***> wrote:

Hi, sorry that I did not record any training logs. But I would suggest having some generated pseudo images for testing. The contrastive loss simply would not go down to a very low value.

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

lorenmt commented 11 months ago

I think you should measure by the validation performance instead of the training loss which does not reflect the true model's performance.