02:45 < ironstark> zoq: So basically those graphs are static images?

02:54 < sumedhghaisas> zoq: Hey Marcus, I know its quite late for feedback... but could you take a quick look at my draft?

09:10 < zoq> ironstark: yes, we used matplotlib to generate all images and some handrafted html template to generate tables etc.

09:12 < ironstark> zoq :thanks for the info

09:59 < Trion> zoq: if I create workers in an agent with their own environments using #pragma omp parallel default(shared), will they share the weights? I want workers to differ in actions but optimize the same weights. Can multiple optimizers running at same time cause conflict with each other?

10:02 < travis-ci> mlpack/mlpack#2235 (master - 4057d14 : Marcus Edel): The build is still failing.

10:02 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/a8363782719b...4057d144eca1

10:02 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/217731083

10:22 < zoq> Trion: Yeah, that can cause conflicts which aren't that easy to solve right out of the box.

10:34 < cult-> does NumChildren() refers to the size of the tree?

10:34 < cult-> how can I measure the size of the model i created or to validate if its big enough?

12:07 < Trion> zoq: ok now I understand A3C's necessity and beauty :-) The workers have their own environment and a copy of model too. :D YAS!!! So we can use a for loop to initialize workers that will accumulate gradients for n steps and apply in order that is marked by Agent's workerNo and only allows a specific worker to make change taking away the conflicts

12:58 < sagarbhathwar> Is there a binomial/multinomial sampling feature in mlpack?

13:06 < sagarbhathwar> Found it in the C++ standard library! Thanks

13:24 < zoq> Trion: I agree it's a really simple and neat idea.

13:24 < zoq> sagarbhathw: Sounds like you figured it out.

13:26 < sagarbhathwar> zoq: Hey Marcus. I know the deadline for GSoC is less than 24hrs from now. Is it too late to submit the proposal?

13:28 < zoq> sagarbhathw: I don't think it's too late, as you said the deadline is April 3.

13:28 < sagarbhathwar> zoq: Maybe you won't be able to review it before submission

13:30 < zoq> sagarbhathw: Yeah, I guess it's unlikely that we have the time to review new drafts and to make comments at this point.

15:17 < kdkw> Is it a good idea to create a PR for minor typos that I have found while reading the source code?

15:21 < zoq> kdkw: Of course: "restoring balance to the universe" -> https://twitter.com/Nick_Craver/status/845963072305057792

15:43 < kdkw> Haha, I will shortly restore balance to the universe :)

15:51 < sagarbhathwar> zoq: I have shared my draft. This is my first time writing a project proposal, so if you feel it's a huge blunder, feel free to let me know :P

15:52 < sagarbhathwar> zoq: That is, if you have time!

16:07 < kdkw> zoq: created the pull request

16:29 < manojbharadhwaj> Hi

16:30 < manojbharadhwaj> Anyone?

16:30 < zoq> manojbharad: Hello, there!

16:31 < zoq> kdkw: great

16:41 < cult-> HoeffdingTree object has constructors only with parameters. It makes it very hard to store the object with constructor initializers. How should I store it in memory?

16:41 < cult-> Whenever I tried to store it as a pointer (at other algorithms) it did produce super strange undefined behaviours

16:41 < cult-> So i assume, that would be a false path this time too

16:42 < cult-> Serialization is an overhead to store it in memory only

16:42 < cult-> Obviously I am missing something in c++ ?

17:09 < manojbharadhwaj> How do I make a submissiom?

17:09 < manojbharadhwaj> Submission *

17:12 < vss> register here https://summerofcode.withgoogle.com/ and then click on mlpack and submit your proposal

17:15 < vivekp> zoq: Just opened a PR for RMSprop update policy.

17:15 < vivekp> Correct me if I'm wrong -- I guess, this was the last remaining optimizer that could be implemented as a SGD update policy?

17:19 < manojbharadhwaj> What is the format to upload a submission ?

17:19 < manojbharadhwaj> Thaks

17:19 < manojbharadhwaj> Thanks*

17:22 < vss> https://github.com/mlpack/mlpack/wiki/Google-Summer-of-Code-Application-Guide

17:23 < vivekp> manojbharadhwaj: It's recommended to upload your final proposal in PDF format (if that's what you meant?)

17:25 < manojbharadhwaj> Thanks

20:06 < kdkw> @zoq, I believe mlpack doesn't currently support deconv, right?

20:41 < zoq> kdkw: Right, there is no deconvolution/transposed convolution layer.

20:54 < kartik__> <zoq> im working with proposal to complete all 4 algorithms in Neural Evolution project in that i liked 2 features to be added. to write and load config file to save and resume saved progress ..

20:55 < kartik__> 2nd to plot the network in real time for which im thinking of using cairographics which is a C++ library ..

20:56 < kartik__> will mlpack allow this additional dependancy

20:56 < kartik__> https://www.cairographics.org/

20:59 < zoq> kartik__: I think we could use the serialization code to load/save/resume the training process. I don't think adding another dependency for a single method is a good idea, but it could be an option for the models repo.

21:02 < kartik__> thanks, i will see it .. so should i drop the idea to plot the network in real time ..?

21:03 < kartik__> <zoq> can u guide me more about the serialization code..

21:08 < zoq> kartik__: You could build a demo an top of the method and use some additional dependencies. That is basically what we do if we train the method on the mario task.

21:08 < zoq> kartik__: The serialization is based on http://www.boost.org/doc/libs/1_63_0/libs/serialization/doc/serialization.html and allows you to save/load defined model parameters (xml, csv). You could take a look at the HMM or NBC code.

22:13 < kartik__> thanks <zoq> that really helped

23:19 < travis-ci> mlpack/mlpack#2243 (master - eca65e4 : Ryan Curtin): The build was fixed.

23:19 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/4057d144eca1...eca65e46a580

23:19 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/217893786