I'm working with accord.net for a while now, btw a greate api! , I'm using multi class svm's with rbf kernel and SMO as a learning algorithm, My dataset consist of 53 labels of 5k samples each, Each sample consist of 400 features scaled using z-scores I'm trying to run a grid search with crossvalidation to tune the hyperparameters (C and gamma), But the performence are not very good ,.. When i use a partial dataset which consist of 200 samples for each label , one 2-folds CV run takes about 30 mins, So a full grid search with 5x5 grid will take about 12.5 hours.. Now, if it was on my full dataset i would be happy with this results but it's on less the 10℅ of it,.. I don't even try to use all of my dataset..,
The complexity range i want to explore is 10^-5 .....10^5 , and the gamma range is 10^-9.....10^3, 5 values for each of the above, Is there anything i can i do to accelerate the proccess?