00:19 < zoq> partobs-mdp: Sorry for the slow reponse, the left side is the expected total reward in the long run, which can be can be estimated from the predicted rewards. Maybe this: http://rllab.readthedocs.io/en/latest/user/implement_algo_basic.html is helpful here? Also, just to be clear you like to implement the REINFORCE model not the fully differential model? I think the implementation of the differentiable model

00:19 < zoq> is easier, but it's up to you.

06:23 < MikeLDN> rcurtin: (regarding the Range Search LNK error in VC) Release 2.0.3 is building OK. VC has /FORCE:MULTIPLE linker flag but it is not working. If I dig some solution I'll write it here...

10:41 < partobs-mdp> zoq: Implemented changes from your review. Can you outline how to crate a PR with all optimizer changes that were made during our previous work? (I understand that I should create a new branch, but what should I do next? I'm slightly confused)

10:45 < zoq> partobs-mdp: You could cherry pick the commits: https://nathanhoad.net/how-to-cherry-pick-changes-with-git; not sure what you mean with "what should I do next", can you eloborate on that?

13:36 < travis-ci> mlpack/mlpack#2890 (master - cea8fe2 : Marcus Edel): The build was fixed.

13:36 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/923df37f9e36...cea8fe2e3b2c

13:36 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/255652299

17:05 < kris1> What layer to use for 0/1 loss in the OutputLayer of ffn

17:14 < zoq> kris1: cross entropy or negative log likelihood loss should work, Konstantin implemented the cross entropy layer here: https://github.com/mlpack/mlpack/pull/1005

17:16 < kris1> I was using the nll but when target(i) = 1; then size_t currentTarget = target(i) - 1; the current Target wraps around so i don;t think nll would work

17:20 < zoq> kris1: ah, right

17:20 < zoq> kris1: however the cross entropy layer should work

17:31 < kris1> Ok i will try that.

17:33 < kris1> One more thing if we do something like a = arma::mat(b.mempt(), b.n_rows, b.n_cols, false, false). Then reassign b = newObject. that would mean that b.n_rows and b.n_cols would also change. Would a change accordingly or not.

17:36 < zoq> kris1: a does not change no, also it's unsafe to do that since c uses the memory from b.

17:38 < kris1> okay…..so how should i share the predictors and responses variable of gan with the generator and discriminator network. Note that both predictors and response change when training disrciminator and generator.

17:41 < zoq> kris1: As long as you don't change the memory pointer or resize the matrix you can use memptr, you could also use a pointer that points to the shared parameter.

18:06 < kris1> mikhail i have added the gan pr. I working on the training right now since it does not work correctly. But the other functions do work. If you have the time please look at tit

18:47 < lozhnikov> kris1: I want to fix the ssRBM implementation first. Nevertheless, I can take a quick look.

18:48 < kris1> yes i will get to the ssRBM tonight. I was working on the Gan from the afternoon. I think cross entropy would fix most of the errors i am getting.

18:48 < kris1> Let’s see.

18:48 < kris1> I will fix the problems you mentioned in the ssRBM PR tonight.

22:22 < kris1> zoq: Could have a look at this https://gist.github.com/kris-singh/fb455fa809634bc8f1afc2872407352d

22:22 < kris1> I am again getting into memory errors ie the program is getting killed.

22:26 < kris1> You could also look at the PR if you want i have used the cross entropy form konstantin

22:30 < zoq> kris1: Can you be a little bit more specific e.g. line number? Just glanced over the code, couldn't see anything right away.

22:30 < zoq> I guess, GanTest (PR) is the test that results in a memory error?

22:32 < kris1> Yes. So well the program just get’s killed without giving line number. But the with valgrind the last line being executed was at gan_impl.hpp:171

22:39 < zoq> kris1: hm, maybe you can step through the program line by line (gdb) to narrow down the error, if not I'll take a closer look at the issue tomorrow.

22:40 < kris1> sorry i did that with lldb and execution stops at 171. I checked the variables there they seemed fine to me.

22:40 < kris1> But i will have another look just in case

22:43 < zoq> okay sounds good, I'll set it on my to do list for tomorrow and get back to you once I know what happened here