error classification (entirely 1st class) with InfoGainLoss layer caffe

Question

according to InfogainLoss layer I have replace SoftMaxWithLoss -> InfoGainLoss from branch https://github.com/shaibagon/caffe/tree/upgrade_infogain (which robustly combines softmax layer and infogain loss layer).

Now all predictions are 1st class only.

Any suppositions?

additional info solver, net, H and log my sover, net, generator H matrix, and log.

I cannot access the gdrive link you posted. can you post the information somewhere more accessible? — Shai, May 29 '16 at 05:15
Hi Alkor, it seems you have (perhaps accidentally) created two accounts; https://stackoverflow.com/users/5982852/alkor and http://stackoverflow.com/users/5982868/alkor. If you wish to merge them, please follow the steps at http://stackoverflow.com/help/merging-accounts — Matt, Jun 11 '16 at 15:14

score 0 · Accepted Answer · edited May 29 '16 at 10:26

0

setting base_lr: 0.001 instead of 0.01 has solve the problem.

edited May 29 '16 at 10:26

Shai

answered May 29 '16 at 09:42

Alkor

1 Answers1