Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

About ranking loss and EuclideanLoss #1

Open
irvingzhang0512 opened this issue Aug 16, 2019 · 0 comments
Open

About ranking loss and EuclideanLoss #1

irvingzhang0512 opened this issue Aug 16, 2019 · 0 comments

Comments

@irvingzhang0512
Copy link

Hi, thank you for your paper and code.

I read the paper and try to understand the model in train_val.prototxt .
As far as i know, final1 & final2 are both ranking functions, and labelr produces labels of {-1, 1}, then the ranking_loss is EuclideanLoss((final1 - final2) * labelr, 0)

my question is:

  1. For me, EuclideanLoss((final1 - final2) * labelr, 0) is quite strange, because labelr seems to be useless in this loss function.
  2. EuclideanLoss((final1 - final2) * labelr, 0) is quite different from the loss function in paper, which i think is like [ max(0, (final1 - final2) * labelr) ]^2.
  3. For both of the above loss function, I think gradient vanishing will occur during training, because both final1 and final2 could be very small(say 1.1e-5, 1.2e-5), and the loss will be very small. Is that the case?

Did i miss something important?
Looking forward to your reply. Thanks again.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

1 participant