00:49 < sumedhghaisas> rcurtin: Hey Ryan... you there? I am running some NTM copy task tests on my machine and they are taking too long... till now I have completed till sequence length 5... NTM performing way better than LSTM for lower sequence length NTM converges in 1 iteration mostly...

00:50 < sumedhghaisas> can I run tests with longer sequence length on the server?

04:03 < rcurtin> sumedhghaisas: you have an account on masterblaster, right? give it a shot there

04:03 < rcurtin> just don't use all the cores :)

10:34 < zoq> kris1: Let me take a look into the issue.

10:36 < kris1> I tried a couple of other things as well …. mainly i changed my default compiler to gcc7 and g++7

10:36 < kris1> but now i am getting boost errors so i am installing the boost with gcc now.

10:39 < kris1> I also tried updating the llvm version now my clang-v gives me Apple LLVM version 8.1.0 (clang-802.0.42)

10:39 < kris1> Which has support for openmpii

10:39 < zoq> Yeah, I think the compiler you use does not support OpenMP which is fine, because OPENMP isn't a requirement. We just have to make sure that if don't use any OpenMP functionality in this case.

10:44 < kris1> So i would have to modify the cmake file right.

10:44 < kris1> Can you guide what changes i should make

10:48 < zoq> The good thing is we don't have to modify the CMake file since it already detects if the compiler supports OpenMP or not. But in parallel SGD we used 'omp_get_thread_num();' which is not a preprocessor macro so the build fails since there is no such function on systems that don't support OpenMP. I will open a PR that fixes the issue soon.

10:51 < zoq> Since the travis and the appveyor build support OpenMP, the build was just fine. I think we could create a new travis job that builds the code without OpenMP support.

10:55 < kris1> How should i proceed in the meantime should i just comment parallel sgd in CMake file in the optimizers

11:01 < zoq> Just give me 15 more minutes.

11:43 < zoq> kris1: https://github.com/mlpack/mlpack/pull/1077 let me know if that fixes the issue for you.

12:54 < kris1> Thanks it worked....

16:36 < zoq> kris1: Just read your blog post, I see the problem with the unit test, I guess what we could do is to use a fixed seed and check if the first outputs are the same as with some precalculated outputs.

17:00 < kris1> But i think fixing the seed of math::Random would effect all test cases… thats why we don’t prefer using seed values

17:01 < kris1> I am trying to come up with a classification test for gan’s also since the discriminator learns the distribution we could test it out on the binary dataset that is skewed but i have to think about it more.

17:02 < kris1> At present i just want to get the GAN working first.

17:03 < zoq> We could just use a fixed seed or fixed parameters for the GAN test. But I agree, let's fix the RBM and GAN first.

22:08 < ironstark> zoq: rcurtin: I was looking at the mlp implementation of dlibml. We have to specify the nodes in the hidden layer. What should we fix it to? Should we ask the user to specify this ? Please let me know your thoughts on this. http://dlib.net/mlp_ex.cpp.html

22:32 < zoq> ironstark: I think we should only test predefined models since there are too many options we would have to support.

22:32 < zoq> ironstark: So I think the only option a user should specify is the name of the model like AlexNet.

22:32 < zoq> ironstark: What do you think, about the idea?