Auxiliary Link Prediction Objective and Entropy Regularization? #8

Activity

1. RexYing commented on Mar 8, 2019

   

   RexYing

   on Mar 8, 2019

   Owner

   More actions

   1. Yes we use NLL instead, since Frobenius norm / L2 loss is harder to train in experiments

   2. I'll add ent later. It's merged from another repo and I didn't put that in.
      It's just a sum of row-wise entropy of the assignment matrix
2. markcheung commented on Mar 9, 2019

   

   markcheung

   on Mar 9, 2019

   Author

   More actions

   Thanks.
3. markcheung

   closed this as completedon Mar 9, 2019
4. usherbob commented on Jun 17, 2019

   

   usherbob

   on Jun 17, 2019

   More actions

   Hi, Rex. Thanks for sharing your code. But I think you should mention NLL loss in your paper. At first, I did not find your diffpool repo, and tried to reproduce your code with L2 loss. It really confused me a lot since the model always went to NaN, then I had to remove the LP loss.
5. usherbob commented on Jun 18, 2019

   

   usherbob

   on Jun 18, 2019

   Last edited by usherbob

   More actions

   Thanks.

   Hi, mark, did you figure it out the calculation of the Link Prediction loss. Could you give me a detailed explanation? Thanks.