Closed JoaoSantinha closed 4 years ago
Hello, I do not see any obvious error, and won't have time to debug myself your code. I would recommend to check that the OWL prox is correct (with some examples where you know what the output should be). Then maybe print intermediate results during optimization to understand when and why it converges to zero. Maybe zero is the solution?
Thanks for your answer. I was able to find a bug (decreasing = False
) through a full implementation, but changing that part has not solve the issue here. I will close and post the solution when I get sometime to investigate this again :)
Hi,
Thanks for really nice package.
I was trying implement the OWL norm but I wanted to check with you if you see something wrong on the implementation and get your help if possible as the weights and intercept are always 0.
where the prox of owl is:
The call I made was: glm.apg(X_train, as.numeric(Y_train)-1, family = "binomial", penalty = "owl", opts = list(lambda1=0.0001, lambda2=0.01))
and I tried with several lambda1 and lambda2. Any other information that you may need just let me know.
Thanks in advance