search

yuvalatzmon / COMET

Code and data for the paper "Learning Sparse Metrics, One Feature at a Time", Y. Atzmon, U. Shalit, G. Chechik, JMLR 2015
3 stars 1 forks source link

hyper-params setting #1

Open HiKyleMorris opened 6 years ago

HiKyleMorris commented 6 years ago

Hi, Yuval

I am trying to solve a multi-class classification problem. The data matrix is very sparse. When using COMET, I find a lot of warnings "Skipped update, origin is not PD". Is this problem related to the setting of hyper parameters ?? How to set these parameters??

Thank you!

yuvalatzmon commented 6 years ago

Hi Peiyan

Thank you for your interest in our work. I need to refresh my memory. I will check it and get back to you.

In general:

This warning is related to the approximation described in appendix E (citing: "In practice, we found ... Therefore, in our experiments we simply skip the update in case its O is not PD."). It indicates that the origin (O) matrix is not PD for the current step, and therefore we skip the current update. In our experiments we found that it does not happen very often and therefore we did not apply the full solution described in appendix E (the full solution is describe at "If, on the other hand, O is not within the PSD cone, we need to...").

Best, Yuval

On Wed, Dec 27, 2017 at 10:39 AM, Peiyan notifications@github.com wrote:

Hi, Yuval

I am trying to solve a multi-class classification problem. The data matrix is very sparse. When using COMET, I find a lot of warnings "Skipped update, origin is not PD". Is this problem related to the setting of hyper parameters ?? How to set these parameters??

Thank you!

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/yuvalatzmon/COMET/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/APwwM1dJCqaEEy_wJOPrhOE352oU4Hy3ks5tEgI8gaJpZM4RNSyB .

HiKyleMorris commented 6 years ago

Hi, Yuval

Thank you for your kindly help!! In my work, there are too many classes, which causes a large constraints (triples). Also, due to the extreme sparsity of features, the algorithm cannot converge.

I am trying to search for parameters in a large space. See if it will be OK?

Best wishes, Peiyan