04:31 < rcurtin> I kept exploring videos of destroying washing machines, but I think my search terminated when I found this: https://www.youtube.com/watch?v=9MDZMr58uEg

10:55 < nishu> hi there

10:55 < nishu> Sir

10:56 < nishu> I'm new to this platform and want to contribute in whatever way I can

10:56 < nishu> I read about your company and would like to comtribute something in it

11:26 < kris> i was implementing a new initialization method for weights of a ann namely sqrt(6/fanin + fanout)

11:26 < kris> how should i get the fanin and fanout values.

12:26 < zoq> kris: I guess fanin is the number of units in the lower layer and fanout is the number of units in the upper layer. So the initialization method needs to know the model structure. I think one way would be to call the method with the model to be initialized.

12:27 < zoq> kris: Something like this: initializeRule.Initialize(parameter, parameter.n_elem, 1, network); and in the initialization method: void Initialize(arma::Mat<eT>& W, const size_t rows, const size_t cols, std::vector<LayerTypes>& network)

12:33 < zoq> To get the input and output size you could use the ParametersVisitor class:

12:33 < zoq> arma::mat parameters;

12:33 < zoq> boost::apply_visitor(ParametersVisitor(std::move(parameters), network[i]);

12:33 < zoq> in = parameters.n_rows;

12:33 < zoq> out = parameters.n_cols;

12:34 < zoq> but I think it might be better to write a new visitor for InputSize and OutputSize.

12:35 < zoq> rcurtin: https://www.youtube.com/watch?v=dq6T5BojXc8 is still my favorite method

12:37 < nish21> zoq: i was thinking about implementing some more optimizers, we already have rmsprop, adam etc. What do you think about adding FTRL?

12:45 < zoq> nish21: Hello, I haven't really looked into FTRL but I remember it's like the online gradient descent method, not sure that's right. If you like to implement the optimizer I'm happy to take a look over the PR.

13:00 < nish21> zoq: okay, i'll work on it and open a pr soon.

13:26 < zoq> nish21: Sounds good.

14:07 < kris> Thanks zoq. Right now implementing the gaussian initialization for weights which has not implemented yet after that i would look into Xavier initialization method.

14:21 < zoq> kris: Nice, let me know if you need any help.

16:10 < rcurtin> zoq: I like that one too, it's very satisfying to watch :)

16:10 < rcurtin> do you have any more comments about the decision tree PR, or should I go ahead and merge that one?

16:15 < zoq> rcurtin: Looks good for me, btw. really nice test suite.

16:16 < rcurtin> yeah, I built the tests as I made each piece, so it was easy to put it all together in the end and know it worked

16:53 < travis-ci> mlpack/mlpack#1828 (master - 3e74746 : Ryan Curtin): The build was broken.

16:53 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/a581cef1508b...3e747462b73e

16:53 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/202956853

18:07 < ozym4nd145_> Hi

18:08 < ozym4nd145_> I am intereseted in Reinforcement Learning project and Augumented RNN Project in GSOC 2017

18:08 < ozym4nd145_> please guide on how to get started on contributing in this org

18:10 < zoq> ozym4nd145_: Hello, have you seen these pages: http://www.mlpack.org/gsoc.html and http://www.mlpack.org/involved.html? Both of those pages should provide good directions to get started.