00:25 < zoq> uzipaz: Also, using the BinaryClassificationLayer layer as output layer, should also decrease the runtime, but I guess, I'll have to add the confidence feature first.

00:26 < uzipaz> zoq: is there much of a difference in implementation? I took a look at both BinaryClassificationLayer and MulticlassClassificationLayer and the only difference was in OutputClass

00:28 < zoq> uzipaz: ah, right, I thought about a different layer

01:17 < travis-ci> mlpack/mlpack#714 (master - ba826b1 : Ryan Curtin): The build was fixed.

01:17 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/f0675d7789b6...ba826b1959a3

01:17 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/121845714

01:31 < uzipaz> zoq: If I use RMSprop as the optimizer, when I call predict on the FFN, the prediction matrix either outputs probability 1 on all cases or 0 on all cases

01:31 < uzipaz> zoq: doesnt make any sense

01:55 < uzipaz> zoq: calling the Train function on FFN with MiniBatchSGD is giving me segmentation fault

06:38 < ranjan123> zoq: you there ?

07:35 < ranjan123> I think I got some serious bug In existing sgd while using sgd to optimize a function for my research work. Could you please confirm whether it is bug or am I making some mistake using it?

11:56 < travis-ci> mlpack/mlpack#715 (master - f4b3464 : Marcus Edel): The build passed.

11:56 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/ba826b1959a3...f4b3464fce6b

11:56 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/121896698

13:04 < uzipaz> ranjan123: I am using SGD with FFN for a classification problem, can you describe the error that you are seeing?

13:52 < uzipaz> i tried using MiniBatchSGD wtih FFN and im getting segmentation fault

13:53 < zoq> uzipaz: I fixed it in: https://github.com/mlpack/mlpack/commit/0e8d776e03b8dbe8e605063b388115cb22b1860d

13:56 < uzipaz> zoq: wow, thanks

13:59 < uzipaz> zoq: when using SGD or MiniBatchSGD, I think we should move the tolerance check outside the currentFunction % numFunctions == 0 check, and also before calling arma::shuffle we need to set a seed for the rng

14:11 < zoq> uzipaz: If we do that, we compare the lastObjective with the current value of the overallObjective parameter and not with the overallObjective over all samples. We have to evaluate the same amount of samples, before we can check the tolerance. If we move the tolerance ouside the numFunctions check, we check the tolerance after each sample.

14:13 < zoq> uzipaz: Why do you think we have to set a seed for the generator?

14:14 < uzipaz> zoq: I tried arma::shuffle on simple matrix/vector and each time I ran the program, I got the same result

14:20 < zoq> uzipaz: So this is only a problem if you run the Train function more than once right? In this case, you could use the RandomSeed function core/math/random.hpp before calling Train again. Maybe it's a good idea to do that inside the optimizer, not sure right now.

14:25 < uzipaz> zoq: I was thinking about this in the loop inside sgd optimize function, where we call arma::shuffle on the visitationOrder after each rotation around the trainingsamples

14:30 < zoq> uzipaz: Not sure what you mean, you mean, we should shuffle the visitationOrder inside the main for loop?

14:31 < uzipaz> zoq: I meant that before we call arma::shuffle on visitationOrder vector on line 83, we should change the seed of rng

14:38 < zoq> uzipaz: ah, I see; If that returns the same order every time this is definitely a bug. Maybe, you can open a issue on github?

14:38 < uzipaz> zoq: sure, will do

14:38 < zoq> uzipaz: Thanks!

16:48 < travis-ci> mlpack/mlpack#717 (master - 592fdcd : Marcus Edel): The build passed.

16:48 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/0e8d776e03b8...592fdcda156a

16:48 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/121927885

17:01 < uzipaz> zoq: If I use MinibatchSGD with batchSize = 1, is it safe to assume that it will behave the same as SGD with other parameters being identical?

17:05 < rcurtin> uzipaz: yes, it will behave the same

17:32 < zoq> uzipaz: Btw. I thought line 83 would always return the same order, but that's not the case, so I'm not sure, there is a reason to set the random seed.

17:35 < uzipaz> zoq: sry if I caused confusion, your right, it does not return the same order, even if using the same seed, the visitationOrder in each rotation will indeed be different, but only if your run SGD as optimizer in the same experiment with identical parameters, we assume identical result each time, because the seed is identical

19:33 < uzipaz> zoq: im running FFN with MiniBatchSGD with batchSize = 50, maxIterations = 100,000. with one hidden layer. Dataset contains 773 features and 842 samples... its been training for 1 hr and 20 mins now and still ongoing... is this normal??