SVM Solver and Testing

[ichitaka]

So i got some questions for the SVM implementation:

1_I implemented a sub-gradient SVM solver (if you haven't found it yet, you're welcome ) and the classic cvxopt optimization of the dual problem. Since we are not allowed to use anything besides numpy, the sub-gradient solver is probably the only way to solve the problem right? Would love to hear any inspiration. Sadly the sub gradient solver needs 34 SVs while the cvxopt solver only needs 17 so i'm not yet really satisfied with this solution.

2_The some constrain never works out for me, even with the cvxopt solution, which kinda confuses me. Any suggestions?

[tanne]

1) First, your are allowed to use optimization and solver libraries for the SVM part! The only numpy constraint is only for the neural network part! So cvxopt is fine here!

2) Some constrain? What do you mean? Dont get that question sorry.

[ichitaka]

Thanks for the answer! When the class started, we were told that we were only allowed to use numpy and matplotlib, that confused me a little now but better to know it later than never i guess.

I meant the sum constrained for a_i*y_i. Maybe my solver messes it up, but it is actually never close to 0. But by concentrating on the cvxopt solution, i might fix this.