01:38 < ironstark> zoq : Thanks for the info.

15:52 < naxalpha> Hi, guys I am back after a long and hectic week

15:52 < Trion> Welcome back :-)

15:53 < naxalpha> Thanks :D

15:53 < naxalpha> I have to discuss #919

15:58 < Trion> zoq: I am planning an Evolution Strategy based agent, but weights in layers are private attributes, how can I add a randn matrix in weights?

16:01 < zoq> Trion: layer.Parameters() returns the weights so layer.Parameters().randn() should work in your case I guess?

16:01 < zoq> naxalpha: Sure go ahead.

16:02 < naxalpha> Is nuget package for executable is really good idea?

16:03 < naxalpha> because as a windows user, I know that I will never use NuGet for installing command-line app

16:03 < naxalpha> instead I would prefer chocho package for that

16:04 < naxalpha> or it would be nicer if i get GUI instead of command-line

16:05 < naxalpha> yes in linux, apt-get is standard to install packages

16:05 < naxalpha> but if we install packages with NuGet, it downloads local package in current directory with extra meta data

16:06 < naxalpha> even if we add package directory to PATH,

16:06 < naxalpha> every person installs package, this directory will go in PATH

16:06 < zoq> We could also provide an chocho package for the executables, I guess the idea is there is the possibility to use nuget.

16:07 < naxalpha> causing lots confusion

16:07 < rcurtin> my opinion is that we should only provide one package; I don't think it's necessary to put it in the user's PATH

16:08 < rcurtin> to say to a user 'you have to install the base library with nuget but everything else with choco'

16:08 < rcurtin> that's a confusing thing, in my opinion that's more confusing that just giving them the executables in the package directory and letting them set their path however they like

16:08 < rcurtin> remember that if you have a command prompt you can just cd to the directory of the mlpack programs and use them from there

16:11 < rcurtin> if you can show me examples of where other software has packaged executables through a different way than nuget but packaged the base library through nuget, I would appreciate it

16:11 < naxalpha> ok fine! but still i am confused why would someone do this as a windows user?

16:11 < rcurtin> why someone would do what?

16:12 < naxalpha> install executables locally

16:12 < rcurtin> I don't mean to sound combative, I am just trying to help come to the best solution here; sorry if it comes off that way

16:13 < Trion> zoq: I was thinking of creating a copy of main model for each worker and then add the noise inside the weights, this method will get me the weights but then how can I make a new model that uses the modified weights? Maybe there should be a SetParameters function in all layers?

16:13 < rcurtin> ok, I see, nuget is typically used just to install a dependency for a visual studio project

16:13 < naxalpha> I am also just discussing alternatives,

16:13 < rcurtin> this is different than what I remember from before

16:14 < naxalpha> mostly windows user like installer (msi) which can be downloaded from mlpack website (most suitable solution)

16:15 < naxalpha> also we can provide downloads for linux (deb) and other platforms etc

16:15 < rcurtin> naxalpha: we will not provide debs, that is the responsibility of the downstream distributions

16:16 < rcurtin> I agree that an msi installer would be the most useful, but we don't have an automated way to set that up

16:16 < naxalpha> I think, I might be able to set that up!

16:16 < rcurtin> this nuget / chocolatey thing is kind of a mess... one is for libraries being used in development of a particular project and another is a systemwide package manager ?

16:16 < naxalpha> yes

16:17 < naxalpha> exactly and also NuGet is mostly used for .Net Libraries but there is room for C/C++ Libraries as well

16:17 < Trion> yeah, windows user prefer msi, everyone likes installing things by just pressing Next Next Next :P

16:17 < naxalpha> :v

16:17 < rcurtin> so I agree now that I read this that if nuget is a development-only tool, there would be no need for including the command-line programs in that (I guess it would not hurt to do so, but it wouldn't be likely to be used in that way)

16:18 < rcurtin> chocolatey seems like the better solution for a user who actually wants the mlpack library and command-line tools installed on their system

16:18 < naxalpha> yup

16:18 < rcurtin> but it also seems like chocolatey can be built on an MSI, suggesting that maybe an MSI would be a first step towards that

16:18 < rcurtin> I'm sorry I misunderstood earlier, nuget has changed from what I remember it being before

16:19 < naxalpha> No problem! I may not be old enough to remember that.

16:20 < zoq> Trion: Not sure I get your question, you are right layer.Parameters() returns the weights you can do the same to modify/set the weights like: layer.Parameters() += arma::randu(...) adds noise or layer.Parameters() = arma::zeros(...) instalize the weights with a zero matrix you can also alter a single parameter: layer.Parameters()(0) = 0.5.

16:20 < rcurtin> at least windows is making some attempt to have some kind of package manager, as long as I can remember windows system administration has always been a disaster due to lack of coherent package management

16:20 < rcurtin> chocolatey might be a step in the right direction...

16:20 < rcurtin> I have to get lunch, I will be back in a while

16:20 < naxalpha> ok

16:20 < rcurtin> thank you for the clarifications about these windows package managers...

16:21 < Trion> zoq: I didn't know assignment to fiunctions was possible in C++ and that even changes the weights!!! :O That's cool!

17:51 < kartik_> hi <zoq> Sry im late..

17:52 < kartik_> should we start with the NEAT implementation ?

17:53 < kartik_> the potential error with large MLPs .. that failed it to complete super mario ..

19:00 < zoq> rcurtin: Do you have any information about what issues they can detect?

19:02 < zoq> kartik_: Not sure what you mean by start with NEAT? If you like we can finish the CMAES method, if you push the latest updates I'll go over the PR.

19:08 < rcurtin> zoq: no idea, we'll see what they come up with

21:38 < ben_______> hey guys, I'm new to this project. What is the preferred method for submitting and getting feedback on pull requests?

21:47 < zoq> ben_______: Hello there, GitHub is the preferred way, the process is straightforward, also http://www.mlpack.org/involved.html might be helpful.