QDA

[iglesias]

I messed a bit the previous pull request ( and my fork in general :S ) with commits that had nothing to do with QDA. I opened a new fork of shogun and have ordered the stuff properly.

Here it goes QDA. There are still a couple of comments in TODO dealing with a couple of doubts I have; they are just about things that probably could be done better but I was not really sure whether it pays off. I am waiting for your feedback :)

[sonney2k]

yeah you should really - precompute as much as is possible. then there is also no need to use the invsqrt approximation. so just store extra variables (or others processed).

[sonney2k]

your patch looks very nice to me - just fix the remaining TODOs and double check that it works (e.g. run it through valgrind / compare results with the scikits implementation) and it is fine to be merged!

Thanks a lot!

[iglesias]

Here it is with the new changes we talked about. A quick summary:

• I have moved the computations in apply that were only dependent of the
  training data to train_machine. This implies a couple of new class members
  and the removal of two class members that were used before in apply but
  now they're not necessary.
• For the serialization issue we discussed in IRC. SGVector* members are
  now SGMatrix and SGMatrix* members, SGNDArray. I have extended a bit
  these classes SGMatrix and SGNDArray with some methods that are useful
  here.

Finally, the serialization of SGNDArray members is pending since you
suggested to leave it this way and think more carefully how this should be done.

Please let me know about your opinion :)