Parallel RF

[Saurabh7]

WIP

[karlnapf]

You can use the copy constructor of SGMatrix

[karlnapf]

So this works?
Whats the gain with increasing number of cores?

[Saurabh7]

Yes, it works.
Although calculating oob error is leading to change in values for test.
I suspected this is due to chaning of order in which we append oob indices (multi thread) to dynamic array. So I made it critical for the thread, but still there are some changes.

outputs: https://gist.github.com/Saurabh7/40889d51c856f73c33337408eb3a7fd1

[karlnapf]

I guess you can force a single thread for the result tests.
Multithread tests can be done on the feature subset shallow copy alone, independently of rf

[Saurabh7]

Ok so few things for this:
currently this Mock test which uses mock labels fails: https://github.com/shogun-toolbox/shogun/blob/develop/tests/unit/multiclass/BaggingMachine_unittest.cc#L35. because this shallow_subset_copy method is not implemented in Labels base class.

Regarding If objects are safe to use in other algorithms, get_feature_matrix() returns a copy if subset are present so thats not a problem.
other access methods like get_feature_vector and set_feature_vector will change the data if written using it.

[Saurabh7]

Also I found the changes in oob error were because, randomly generated indices in multithreaded version keep chaning even thought I set seed in sg_rand.

[karlnapf]

And merge :)

[karlnapf]

ah crap, it wasnt squashed ;)
Can you always make sure you squash your commits before asking to merge them?
I know I should check, but there is a lot of stuff going on atm

[Saurabh7]

@karlnapf ah ok sorry :) Next time I will put an explicit comment when final.