00:23 < zoq> ShikharJ: N need to work over the weekend, the debugging process is the hardest part, and I think we are on track.

00:25 < zoq> haritha1313: One way is to use sequential instead of the FFN class, which should work with the merge layer. Another way is to write another class, this is what we do for the FAN code (kinda of a wrapper around two FFN's).

00:28 < zoq> ShikharJ: I think the network isn't initialized correctly at the point where the Gradient function is called.

00:34 < zoq> ShikharJ: If we call the gradient method of the FFN class, the method runs the Evaluate function (Evaluate(parameters, begin, batchSize, false);), which requires that predictors is set. In this case it looks like it is set, but the number of cols (numFunctions) is 1 instead of 500.

00:35 < zoq> ShikharJ: Let's see if the first issue is fixed in the latest version.

00:37 < zoq> haritha1313: I think #1355 is close to be merged, so perhaps we should focus on the Pr for now before moving forward. Let me know what you think.

04:52 < haritha1313> zoq: Sure. Actually a few tests are failing with errors different from earlier. Since it was taking time to solve I thought I'll have these week's work go in parallel.

04:54 < haritha1313_> I'll try my best to debug them and push the changes today itself.

06:27 < zoq> haritha1313: Ah, I see, I'll take a closer look into the failing tests as well.

06:29 < ShikharJ> zoq: I'll take a closer look and reach back if I face any issues.

06:33 < zoq> ShikharJ: It looks like 'generator.Gradient(generator.parameter, noise.n_cols, gradientGenerator, batchSize);' isn't correct. It will call 'Forward(std::move(predictors.cols(begin, begin + batchSize - 1)));'

06:33 < zoq> but predictors.n_cols is 1 and begin is also 1

06:33 < zoq> but it should be zero

06:34 < zoq> still trying to figure out what is wrong

06:54 < zoq> okay turns out that predictors inside the FFN class isn't set.

07:01 < ShikharJ> zoq: I must mention, I didn't change the currentInput for the generator inside Gradient from noise to predictors

07:02 < ShikharJ> It didn't make sense to me at that moment so I had skipped it.

08:08 < travis-ci> ShikharJ/mlpack#160 (GAN - 13494c7 : Shikhar Jaiswal): The build has errored.

08:08 < travis-ci> Change view : https://github.com/ShikharJ/mlpack/compare/291564cf8a70...13494c7c543a

08:08 < travis-ci> Build details : https://travis-ci.org/ShikharJ/mlpack/builds/384015117

10:01 < jenkins-mlpack> Project docker mlpack nightly build build #330: UNSTABLE in 2 hr 47 min: http://masterblaster.mlpack.org/job/docker%20mlpack%20nightly%20build/330/

15:31 < zoq> ShikharJ: inside the Gradient function we already call: generator.Forward(std::move(noise)); you are right, so in the best situation we don't rerun the Forward step inside the Gradient call of the FFN class, but at the moment we do, so we would have to set the predictors once more.

21:00 < ShikharJ> zoq: That makes sense to me. I have updated the changes.

23:30 < ShikharJ> zoq: For the case that you mentioned above, shouldn't instead of doing (currentInput = predictors;) we should set the (gradient.Predictors() = predictors) or something like that directly?

23:33 < ShikharJ> zoq: Ah, nevermind, I just saw that repeated your suggestions above.