04:21 < ironstark> zoq: rcurtin: I have converted the google doc to a blog post. I have to submit the final work within 35 hrs. Please take a look at it : http://mlpack.org/gsocblog/benchmarking-mlpack-libraries.html

08:08 < kris___> lozhnikov: The cnn actually takes 8 min per iteration on my system.

08:08 < kris___> ./gan_test2.o -i train7.txt -o output.txt -e 100 -m 2000 -b 50 -x 5 -N 100 -r 0.003 -v

08:08 < kris___> I am running this setup

08:09 < kris___> so it is not possible for the system to converge at all.

08:15 < lozhnikov> 8 min per iteration??? did you change anything in the test?

08:16 < kris___> just CrossEntropyLogit to CrossEntropy + Sigmoid

08:16 < kris___> with GaussianIntialisation to (0, 0.02)

08:16 < lozhnikov> on my system one epoch takes about 30 mins

08:17 < kris___> I have 8 gb system.

08:17 < kris___> 1.6Ghz processor

08:18 < lozhnikov> well, I ran the test in the morning with the following parameters -e 40 -m 2000 -x 300 -N 100 -r 0.003 -v

08:19 < lozhnikov> I hope the test will finish in the evening

08:19 < kris___> Okay lets see.

09:53 < zoq> ironstark: Looks really good, I'll see if I can fix the table rendering later today.

11:26 < ironstark> zoq: Thanks a lot :)

11:27 < ironstark> Should I post the link on final submission blog and complete it?

11:28 < zoq> ironstark: Sure, I think it's ready.

11:29 < ironstark> zoq: Thanks :)

12:46 < kris___> lozhnikov: I have stopped the test for conv gan

12:46 < kris___> I will work on cifar dataset with ssRBM

12:47 < kris___> And make the resize layer pass the build

12:47 < kris___> Is that okay with you

12:54 < lozhnikov> kris__: looks reasonable to me

13:17 < zoq> ironstark: Just used a HTML table for now.

13:36 < ironstark> zoq: It looks great.. Thanks for doing this. :)

14:19 < rcurtin> ironstark: thanks for the nice writeup, I enjoyed reading it

14:19 < rcurtin> if you open a PR for Shark or MachineLearning.jl, I will be happy to review it and incorporate it :)

14:38 < ironstark> rcurtin: Thanks , I'll open up a PR soon.

14:39 < rcurtin> ironstark: if you like, the next thing to do with a benchmarking system is actually do the benchmarking of the different machine learning tasks and find a nice way to display the results

14:40 < rcurtin> if you had fun working on your project and want to keep playing with the system in the future, you are definitely welcome to, and maybe doing some specific benchmarking might be enjoyable :)

15:27 < rcurtin> sumedhghaisas: serialization changes to the NTM look good to me, thanks for handling that. have you had a chance to try the NTM on some of the benchmarking tasks in the paper? I am excited to see the results :)

15:31 < kris___> lozhnikov: Build for the resize layer and crossentropywithlogit are now passing.

15:31 < kris___> Let me know if there are any more comments.

15:32 < lozhnikov> kris__: sounds good. I'll look through the code in the evening

15:50 < sumedhghaisas> rcurtin: Hey Ryan... yeah I will send the code to 'models' repo. I made sure that the performance is better than LSTM, GRU

15:50 < sumedhghaisas> although the parameters need to be tweaked more for optimum performnce

16:54 < rcurtin> sumedhghaisas: yeah, I am not surprised it needs some tweaking, but great to hear that the performance is better

18:47 < kris___> lozhnikov:

18:47 < kris___> Did the conv gan converge yet...

19:40 < lozhnikov> kris___: the test has been finished. But the results don't look good. I have to look through the oreilly example more detailed maybe we overlooked something

19:41 < kris___> Is it just random noise...again

19:43 < lozhnikov> no, I got black samples