After a weekend of debugging I can proudly say that I am able to run toy examples using my implementation of SO-SVM for the multiclass classification case of use \o/
Just for the record, and for the fun of its discovery, I am going to write about the last bug I fixed in my code. As it is commented in the previous posts, part of the project deals with solving an optimization problem, a QP in particular. The fact is that I could run the code just fine, no more failing ASSERTS popping up nor segmentation faults. However, something weird was going on with the solution. The weight vector was zero for every problem instance. After double checking with MATLAB’s function quadprog that the correct solution for the problem was effectively non-trivial, I tried adding some lines here and there to ensure that the parameters given to MOSEK’s were the correct ones. All the matrices and vectors were ok. After some more lines to print out the bounds of the problem I found it. The variables of the optimization problem were all fixed to zero!
At the beginning I just assumed that if no bound is given for a variable, then MOSEK would consider that this one is free; i.e. it may take values within . However, this is false. If no bound is given, then the reality is that MOSEK fixes the value of the variable to zero. I still don’t get the point of including variables to the optimization vector whose values are fixed… but that’s another story.
For the next time I will try to remember to ensure that the assumptions I make are correct or, even better, to RTFM
My next objective is to try out the application with bigger examples. Unfortunately my MOSEK license file just allows me to solve problems with up to 300 variables or constraints, which currently prevents me from doing tests with training data bigger than 17 bidimensional training examples!