08:59 < partobs-mdp> zoq: rcurtin: Finally got the non-zero parameters in HAM, but the unit test is still messed up. It seems like I've done some armadillo magic wrong, but could you take a look into the issue?

08:59 < partobs-mdp> The code in the latest commit

09:00 < partobs-mdp> One of the problems is that the JOIN doesn't get valid parameters and emits some negative values (despite having weight matrix consisting of positive elements)

15:10 < kris__> https://usercontent.irccloud-cdn.com/file/FFxWD9Es/gan_keras.cpp

15:10 < kris__> Hi lozhnikov could you run this file for me.

15:10 < kris__> These are the parameters ./gan_keras.o -i ./train4.txt -o ./output.txt -m 1000 -e 1000 -r 0.001 -b 100 -g 2 -N 100 -G 1024 -D 1024 -t -1 -s -v

15:15 < lozhnikov> kris__: I have to do another work right now, I'll run your code in an hour.

16:35 < Guest75930> hi. i have just working on open-source projects.

16:35 < Guest75930> How do i get in touch with this organization

16:38 < zoq> Guest75930: Hello there, http://www.mlpack.org/involved.html and http://www.mlpack.org/gsoc.html are probably helpful.

16:38 < zoq> Guest75930: We are always open for new contributions, let us know if we should clarify anything.

16:41 < Guest75930> thank you. will do

18:06 < kris1> lozhnikov: I think now the resize layer is fixed. You can have a look.

19:04 < kris1> Lozhnikov: I was think that we should add a parameter for noise size in Gan.hpp.

19:05 < kris1> Right now we have only one noise sample. On each batch we train the Gan on batchSize * trainData(0, batchSize) + noise*BatchSize.

19:07 < kris1> or we are training on noise very less. So if the batchSize is small this would not be a big matter. But if the batchSize = 100. the diffrence is pretty huge. I think this the reason the gradients are not big in out case.

19:07 < kris1> What do you think ?

20:45 < kris1> Also rather than having an extra parameter i was thinking we always set the noiseSize = batchSize so the now the predictor would (trainData.n_rows + batchSize, trainData.n_cols)