verne.freenode.net changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
kaushik_ has joined #mlpack
govg has quit [Ping timeout: 246 seconds]
< rcurtin>
zoq: I refactored the SGD-like optimizers, maybe the patch would be useful for you too---
< rcurtin>
I put it in github.com/rcurtin/mlpack on the 'batch' branch, commits f29a9ddb fddae469 b3f9a456
< rcurtin>
the logistic regression function changes need a lot of help, I am now debugging those; but at the very least the optimizer infrastructure seems to be ok in those patches I pushed
< zoq>
rcurtin: Where does the initial batchSize (32) come from?
< rcurtin>
I haven't tuned it, just a guess
< rcurtin>
there are some bugs I am fixing now, I will push a fix shortly
< kaushik_>
rcurtin: cli_test doesn't compile now and i get errors as above ^
< rcurtin>
kaushik_: comment out cli_test.cpp, cli_bindings_test.cpp, and python_bindings_test.cpp
< kaushik_>
ok
< rcurtin>
these are the errors that I will fix once you finish the rest of the conversion; I can't fix them until everything uses serialize() instead of Serialize() though
< rcurtin>
or, I could fix them now, but it would be a huge amount of pointless work :) much easier to do after everything is changed
< kaushik_>
rcurtin: can you see the latest changes in methods/hoeffding_trees from PR. I am probably doing something wrong there.
< rcurtin>
I have to feed my cat first, she is not going to stop making noise until I do :)
< zoq>
Fritz? Can't remember the name of the other cat.
< rcurtin>
ha, yeah, the other one is named Drusilla
< zoq>
ah, right
< rcurtin>
kaushik_: ok, I see, the problem comes from line 767 in hoeffding_tree_impl.hpp:
< rcurtin>
ar & BOOST_SERIALIZATION_NVP(numericSplits);
< rcurtin>
I think that because we are now serializing a vector, boost::serialization will call vector.resize(), which needs access to a default constructor
< rcurtin>
I guess that in this case, we can modify the numeric and categorical split types to have a default value for the number of classes
< rcurtin>
zoq: I pushed some more fixes to the 'rcurtin/batch' branch, now logistic regression and NCA work correctly
< kaushik_>
rcurtin: ah ok I see.
< kaushik_>
will do that by tomorrow. Its 4 AM and I need to sleep :D