09:07 < kris1> lozhnikov: hi i have updated the gan PR could you have look.

09:08 < kris1> With low no of iterations i am getting very bad results but i am. testing with high number of iterations(takes lot of time) now lets see how it performs

09:09 < kris1> zoq: if you have the time could you also have a look particularly at the evaluate function and gradient function of GAN. That would be very helpful.

09:25 < lozhnikov> kris1: what about this comment https://github.com/mlpack/mlpack/pull/1066#discussion_r128708271 ?

09:37 < kris1> lozhnikov i actually reset parameters first and then do generator.Parameters()

09:37 < kris1> = arma::mat()

09:38 < kris1> but it works now….

09:38 < lozhnikov> did you test that with valgrind?

09:39 < lozhnikov> I think this is incorrect since the layers still use the previous pointer

09:39 < kris1> you mean for the memory leak.

09:40 < lozhnikov> No, I mean invalid pointer

09:40 < lozhnikov> And I've got the second comment: How do you think, is it possible to move the code of Train() to the Gradient() function

09:40 < lozhnikov> ?

09:42 < kris1> Can you please elaborate. I don’t understand the comment.

09:43 < lozhnikov> the first or the second?

09:43 < kris1> second

09:44 < lozhnikov> I mean is it possible to refactor the gradient function in such a way that Train() contains only "optimizer.Optimize()"?

09:45 < kris1> for 1 i think it is to check i will the where the generator.parameters are pointing to usning memptr if it is the same as parameters.memptr() that would be sufficient i guess.

09:46 < kris1> I would have to think about the refactoring. Intially i was thinking of creating 2 seprate function trainGenerator and trainDiscriminator and then calling them from the train function. But just having optimizer.Optimize would be require some thinking. I will see.

09:47 < kris1> the result just came for 1000 iterations for gan. Will have to see where i am going wrong.

09:48 < kris1> lozhnikov could you comment on if the train and gradient function are correctly implemented.

09:49 < lozhnikov> regarding the first comment: pointers are different since you use the following:

09:49 < lozhnikov> generator.Parameters() = arma::mat(parameter.memptr().....

09:51 < lozhnikov> okay, I'll take a quick look now. I'll look through the code in detail as soon as I finish with the ssRBM

09:51 < kris1> Okay sure……….. i will check comment 1

10:08 < lozhnikov> kris1: It seems the formulas of gradients do not match the formulas in the GAN paper

10:08 < lozhnikov> see https://arxiv.org/pdf/1406.2661.pdf page 4

10:14 < kris1> how do you mean…. for the gradient calculations of discriminator i just call the discriminator.Gradient().

10:15 < kris1> for generator i am passing the error from discriminator.network.front() to the genertor.error() and then i am bascically calling the generator of the generator.

10:16 < kris1> i have taken the fakeLabels to 0 and real labels to 1 btw for disriminator calculation and fakelabels = 1 for generator calculation.

10:27 < lozhnikov> okay, It seems I'have understood that. Right now I haven't got any comment except

10:27 < lozhnikov> Optimizer.MaxIterations() = numFunctions;

10:28 < lozhnikov> I think you should replace that by "Optimizer.MaxIterations() = k"

10:53 < ironstark> rcurtin: zoq: I wanted to discuss regarding the new library we should benchmark. Can we install R and benchmark it? I think it would be good if we could have a benchmarking system that could compare the performance of R against Python. Please let me know what you think

12:09 < zoq> ironstark: We can install R, yes, however my understanding is limited, so I can't provide much help here; I think rcurtin suggested dlib-ml a while back, i guess that could also be an option. Let me know what you think. I'm sure we can find something everyone agrees on.

12:21 < ironstark> zoq: i can work on dlib-ml too but I dont have much experience in that. I can first work on dlib-ml and then R. Please let me know your thoughts on this.

12:44 < zoq> ironstark: Let's wait for rcurtin's thoughts on this. I like both ideas.

18:34 < kris1> i fixed the reset function for the GAN and pushed the changes. Regarding the refactoring of training i am still working on it.

18:34 < kris1> lozhnikov

18:39 < kris1> thought the samples produced still are pretty random it seems. i will work on that first and then refactor the train function.

18:59 < lozhnikov> kris1: I looked through the fix. I think there is an error.

18:59 < lozhnikov> generator.Parameters() = arma::mat(parameter.memptr(),.....

18:59 < lozhnikov> discriminator.Parameters() = arma::mat(parameter.memptr(),....

18:59 < lozhnikov> The pointers should be different.

19:01 < kris1> Aaah right discriminator.Parameters() = arma::mat(parameter.memptr() + weightsGenerator(),….) thats what i meant to do .