06:10 < Habib> Hello, anybody there?!

07:43 < zoq> SinghKislay: Hey.

07:44 < zoq> Habib: Hello.

07:46 < zoq> Suryo: Thanks for the update, will take a look at the PR and comment there.

07:47 < zoq> favre49: See my comment to xain (IRC).

07:55 < zoq> sreenik: what you could also take the norm for each col; take a look at the conv net test for an example.

08:15 < mulx10> Hello

08:15 < mulx10> I need help regarding Memory checks in my PR.

08:17 < mulx10> I ran valgrind with same arguments. It works on my PC, however it fails in Jenkins.

08:17 < mulx10> https://github.com/mlpack/mlpack/pull/1756

08:57 < Viserion> expected constructor, destructor, or type conversion before ‘(’ token mlpack::data::Load("/home/ans.csv", data, true); // The dataset itself.

08:57 < Viserion> how to resolve this?

09:47 < sreenik> Viserion: Are you running the code globally, i.e., not inside any function?

10:03 < favre49> zoq: Thanks, CoDeepNEAT looks really cool. I'll come back to you with some ideas soon.

10:22 < sreenik> zoq: After norm, I increased the step size x10. That made the validation accuracy converge at around 97%. Normalising along the columns also give more or less the same result. Layer or node addition or dropout also doesn't increase it further. I don't think accuracy can be increased further without using CNNs. So I am trying to optimize the CNN version which originally has ~80% accuracy :)

10:31 < Subrajaa> I am a student from India. I would like to know whom to contact in the organisation so that I can interact with them. I am a beginner and would like to contribute sincerely. Please help.

12:49 < sonu628> hey everyone

12:50 < sonu628> when i compiled and ran programs in folder where mlpack was installed

12:51 < sonu628> it did successfully, however when tried in other i got [FATAL] error

12:52 < sonu628> i thought it must issue with armadillo so installed it with its source code

12:53 < sonu628> now i got this error

12:53 < sonu628> usr/bin/ld: warning: libarmadillo.so.8, needed by /usr/lib/gcc/x86_64-linux-gnu/7/../../../x86_64-linux-gnu/libmlpack.so, may conflict with libarmadillo.so.9

12:53 < sonu628> any suggestions what should i do ?

13:17 < ShikharJ> sonu628: I guess this conflict is due to multiple versions of armadillo being installed. I think the better thing to do would be to remove one of the installations, and then try.

14:08 < sonu628> ShikharJ : yes you were right it was a colllision of armadillo. I have removed all versions armadillo and reinstalled through apt-get and again ran cmake and build, and unfortunately got this error

14:08 < sonu628> warning: libhdf5.so.101, needed by /usr/lib/x86_64-linux-gnu/libarmadillo.so, not found (try using -rpath or -rpath-link)

14:09 < sonu628> after this there were many undefined references of what i think of components needed in libhdf5.so.101

14:09 < sonu628> on doing locate libhdf5.so.101 i got addresses installed in anaconda

14:10 < sonu628> i would be grateful if anybody suggest any solution ?

14:13 < rcurtin> link your program against hdf5 with -lhdf5?

14:13 < manish__> Hi! i have newly joined this chat session. Are all here gsoc2019 students??

14:29 < sonu628> rcurtin : the error comes when try the make command after cmake

14:31 < sonu628> i dont think make -lhdf5 would work

14:45 < rcurtin> sonu628: the CMake configuration should be automatically linking against all dependencies of Armadillo

14:45 < rcurtin> so it's strange that it isn't

14:45 < rcurtin> but you say you have hdf5 installed via anaconda and not the system package manager?

14:47 < sreenik> rcurtin: Today I was setting up mlpack on a gcp instance. I have run into the same problem of hdf5, but this didn't happen in my machine at home.

14:48 < sreenik> In the gcp instance I have hdf5 installed via the package manager apt-get

14:54 < sreenik> In the cmakefile, at the very end there is a something like (autoDetect = false) as far as I remember. If I can get a fix I will post an update here

14:54 < rcurtin> yeah, if you have built Armadillo by hand you can configure with 'cmake -D DETECT_HDF5=OFF .' instead of './configure' (which is just a wrapper script for cmake)

15:39 < sreenik> I think setting setting DETECT_HDF5=OFF while building armadillo has worked. Mlpack is currently in the build process and hasn't thrown any error till now :D

16:30 < sreenik> Built successfully :)

16:32 < shubhangi> Hello Everyone,

16:34 < shubhangi> I wanted to contribute to Application of ANN Algorithms Implemented in mlpack.

16:34 < shubhangi> I started to build mlpack.

16:39 < shubhangi> I started to build mlpack on ubuntu machine 64 bit and it got stuck on running the `make -j4` command

16:41 < shubhangi> I ran it on CPU and it is not processing after 77% where it is building neural network. So, is there any issue with my machine or should I shift to GPU?

16:42 < rcurtin> shubhangi: hi there, is your system swapping?

16:43 < rcurtin> if it is running out of ram you can try with make -j2 instead

16:43 < rcurtin> and that may help

16:47 < shubhangi> At first I tried with -j4 then it stopped, then I did with j and still it is not proceeding further

17:06 < riaash04> zoq: In exploring how NEAT could be used to optimise arbitrary functions, I went through the implementation of CNE since it also does Neuroevolution. So for optimising arbitrary functions CNE considers the parameters as weights and optimises accordingly while evaluating based on the Evaluate method (basically GA).

17:06 < riaash04> But we can't do this with NEAT

17:07 < riaash04> Since NEAT also changes topology so extra wights will get added

17:08 < riaash04> So for using NEAT as an optimiser for arbitrary functions we could consider output values of neural networks as values to evaluate the functions (simliar to sequential tasks).

17:09 < riaash04> The input to the networks would be the initial values of the parameters.

17:10 < riaash04> Since NEAT is good at optimising NNs, this could be a way. (Maybe I am just stating the obvious or am completely out of line)

17:12 < riaash04> This is somewhat similar to what they have experimented in "Is Meta-EA a viable optimisation method"

17:29 < riaash04> Although, this way of optimisation would not be very useful. We would still be able to use it as an optimiser for sequential tasks ans meta-optimiser.

17:55 < favre49> riaash04: The problem i see with the method you described is that it seems wasteful to use something like NEAT in that way when DE or CNE would suffice.

17:59 < rajiv_> zoq: I have sent you an updated proposal idea... Please let me know what you think :)

18:02 < riaash04> Yes that is true. This is just a way of using NEAT as an arbitrary function optimiser along with sequential task optimisation and meta-optimiser without having to add specialised methods.

18:04 < riaash04> favre49: sorry forgot to mention you in the previous message

18:11 < favre49> riaash04: You're right. To me, NeuroEvolutionary Meta-Optimizers seem like the most apt way to do this. There are problems with it though.

18:14 < favre49> For one, we need to give it a substrate, which would be problem specific. Also, the author stated it did not perform well on all functions.

18:14 < favre49> So, perhaps to revise my original statement, it doesn't seem as apt xD

19:09 < Viserion> After compiling by this command: g++ knn_example.cpp -o knn_example -O3 -std=c++11 -larmadillo -lmlpack -lboost_serialization i stll get the following error:tmp/ccQPHIsH.o: In function `main': knn_example.cpp:(.text.startup+0xd5): undefined reference to `bool mlpack::data::Load<double>(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, arma::Mat<double>&, bool, bool)' collect2: error: ld returned

19:10 < rcurtin> Viserion: are you linking against the correct libmlpack.so? what have you tried to debug it?

19:13 < Viserion> I builded mlpack from source and then used the above command. Honestly I do not know how to check weather I am linking to correct libmlpack.so . Can you help me out?

19:15 < Viserion> My c++ file : #include <./mlpack/core.hpp> int main() { arma::mat data; data::Load("/home/viserion/ans.csv", data, true); }

20:07 < rcurtin> Viserion: sorry, I stepped out for a while. did you install mlpack from source?

20:07 < rcurtin> and are there two versions of mlpack installed on your system?

20:08 < rcurtin> you've probably seen this page--- http://mlpack.org/docs/mlpack-3.0.4/doxygen/build.html

20:09 < rcurtin> in the "Simple Linux build instructions" section, it says "On many Linux systems, mlpack will install by default to /usr/local/lib and you may need to set the LD_LIBRARY_PATH environment variable"

20:09 < rcurtin> that also means that if you are linking against mlpack, you may need to specify -L/usr/local/lib on the compiler command line, so that gcc will link against /usr/local/lib/libmlpack.so

20:09 < rcurtin> (which is the one you built and installed)

20:18 < zoq> riaash04: You are on the right track, I agree that CNE might be faster at least in some cases; but that depends on the initial topology, for more complex tasks NEAT should be able to find a solution faster. As Ryan pointed out, usually a new optimizer is tested on a single task/random seed, which shows some advantages over method x, but there is no guarantee that it will work across different task/seeds. So

20:18 < zoq> it might be useful to run some experiments at the end to get some more insights.

20:27 < riaash04> zoq: Ok, thanks. So I will build on this more. :) Also, I was thinking CoDeepNEAT could also be part of project. Like a meta-optimiser application of NEAT to optimise DNNs and it would be able to utilise NEAT's population and speciation (as it uses the same process).

20:34 < zoq> riaash04: It could be part of the project, for now the focus is on NEAT including testing (big point) and documentation.

20:56 < donfreecss> hi everyone,

20:57 < donfreecss> i'm looking for an example of cmakefile that can be used to build c++ projects that uses mlpack, i never used cmake before unfortunately

20:58 < donfreecss> i've built mlpack successfully on my machine but working with it by command-line based compilation .. i.e gcc -lmlpack -l.. -l.. etc

20:59 < zoq> donfreecss: You could checkout the models repo (https://github.com/mlpack/models) which uses cmake to find mlpack.

21:00 < zoq> donfreecss: https://github.com/zoq/nes is another project, the important file here is https://github.com/zoq/nes/blob/master/CMake/FindMlpack.cmake

21:05 < donfreecss> got a quick look and didn't get much :) think i need to read more about cmake and come back and go through these examples

21:05 < donfreecss> Thank you

21:05 < zoq> sounds good, let me know if I should clarifyy anything

21:08 < rcurtin> donfreecss_: if you're just writing simple applications that link against mlpack, CMake may be overkill

21:09 < rcurtin> but if you're trying to write something that you hope to distribute to other people's systems that they should be able to configure and install, then CMake is probably the right tool for that

21:09 < rcurtin> the key file in the models/ repo is probably going to be CMake/FindMLPACK.cmake and CMake/FindEnsmallen.cmake, which can be used to help find those libraries with CMake

21:09 < rcurtin> oops, "key files". also CMake/FindArmadillo.cmake and related files can be really helpful too