04:51 < partobs-mdp> zoq: rcurtin: Got back to the computer, re-implemented Reset() as zoq mentioned. Now it crashes at compile time, because for some reason the FFN<>.Model() method disappeared. How to fix that?

05:03 < partobs-mdp> (update) Weird stuff: merging ffn*.hpp files from upstream into my branch resolved the problem

05:03 < partobs-mdp> Now brushing up the unit test and pushing it

06:13 < lozhnikov> kris__: I don't think that separate training of the generator and discriminator network could improve the results since the optimizer updetes weights independently, the gradient of the generator doesn't depend on the gradient of the discriminator. Actually, our approach doesn't differ significantly. We calculate the gradient of the generator when the weights of the discriminator are not updated yet. That's the only

06:13 < lozhnikov> difference. So, I think our approach should give the same results.

06:14 < lozhnikov> I guess some parameters are incorrect

06:15 < lozhnikov> Did you try to vary the batch size, the discriminator (generator) hidden layer size, the noise size, the network architecture?

13:03 < kris__> Hey zoq: I need a little help...with resize layer.... when i am addding it to the layer_types.hpp. I am gettting a strange error could you help.

13:05 < rcurtin> partobs-mdp: glad you got it worked out; I have some meeting today but I can help out with anything later int he day if needed

13:05 < rcurtin> *later in the day

13:09 < lozhnikov> kris__: Could you describe the issue?

13:10 < kris__> https://www.irccloud.com/pastebin/yrb6fLz8/Error%20message

13:12 < lozhnikov> I mean could you upload the code?

13:12 < kris__> okay sure...

13:22 < kris__> Uploaded the code the error seems to be in layer_types.hpp file.

13:37 < kris__> Lozhnikov: Were you able to reproduce the error....

13:39 < lozhnikov> kris__: remove ',' from layer_types.hpp:120

13:41 < kris__> Ahhh that gives another error for me.

13:42 < kris__> the undefined symbols error that it was giving earlier.

13:58 < lozhnikov> kris__: try to fix incorrect #ifndef at resize_impl.hpp:13

14:02 < kris__> Ahhhh done....

14:03 < kris__> Could you explain the test you were talking about. For resize layer i wrote different test if passed. But i don't think its of much use.

14:06 < lozhnikov> Try to resize a linear function, (for example x + y). The result should be the same (linear interpolation is exact for linear functions)

14:14 < kris__> is the bilinear interpolation distributive over addition. ie should resize(x) + resize(y) == x + y. Scaling factor = 1.

14:23 < lozhnikov> kris__: yeah, I think that's correct

15:50 < kris__> I updated the PR. It took time because i was thinking of the same test for backward pass but found that simple test would not work for the backward pass / un zooming.

15:51 < kris__> we have treat the scale = 1 as special case for backward pass.

16:32 < crater_kamath> Hey guys..I want to contribute to mlpack

16:32 < crater_kamath> I am a beginner and have submitted a patch to mozilla

16:33 < crater_kamath> Can anyone help me in getting started?

16:34 < crater_kamath> and please tell the ML concepts that I need to learn cause machine learning is new to me

16:49 < zoq> crater_kama: Hello and welcome; have you seen http://www.mlpack.org/involved.html or http://www.mlpack.org/gsoc.html

16:49 < zoq> crater_kama: The necessary knowledge highly depends on what you like to do, but as long as you are willing to learn the necessary concepts you are good to go, so some knowledge and is often sufficient. Also, mlpack uses lots of different C++ paradigms including a lot of template metaprogramming, the gsoc page includes some helpful reference that you should checkout if you dig into the code.

16:54 < zoq> partobs-mdp: Sorry for the slow response, sounds like you fixed the issue?

16:54 < zoq> kris__: Build issue solved?

17:01 < kris__> Yes it's done now.

17:02 < zoq> great

18:02 < kris__> lozhnikov: So after changing hte some parameters i think the 1d gaussian example is working now.....

18:03 < kris__> https://usercontent.irccloud-cdn.com/file/PMWccWqr/output_Jmd1ZD.gif

18:04 < kris__> Here are the result i did not set the y scale since the image gets cropped.

18:04 < kris__> ./gan_test.o -o gaussian1_output.txt -e 1 -b 8 -r 0.001 -g 2 -v here are some of the parameters i tried.

18:05 < kris__> I changed the epoch from 1 to 20

18:05 < kris__> but i saw that later the learning doesn't take place.

18:05 < lozhnikov> sounds good

18:06 < lozhnikov> were you able to obtain correct parameters (and the layer structure) for the Digit dataset?

18:07 < kris__> No the parameters are harder to tune for digits.

18:07 < kris__> I would like try the keras example now.

18:07 < kris__> Since i have certain guide for parameters there.

18:08 < kris__> Otherwise the orilley example. I think the resize layer can be merged with few changes let me know what you think?

18:10 < lozhnikov> I think we should implement the test for a linear function first. Actually, you always can merge that PR into the GAN PR

18:11 < kris__> I did implement the test for the linear function. Forward pass alone in the latest commit.

18:12 < kris__> There is no similar test for the backward pass.

18:12 < lozhnikov> ok, I'll think about that

18:14 < kris__> Also I tried changing some parameters for the ssRBM test. Either the test fails or the time difference is not that big.

18:15 < lozhnikov> I know. The test is slow since the ssRBM is ten times slower than the binary RBM. I didn't dig into that yet

19:14 < rcurtin> zoq: I think that FFN::Serialize() should also serialize the network type itself; what do you think? this would allow deserializing any FFN<> and then the structure of the network would be loaded too

19:14 < rcurtin> I am working on a patch for it (because I at least need the support for other things, even if you'd like to leave FFN::Serialize() as-is)

19:37 < zoq> rcurtin: Sounds like a great idea.

20:57 < rcurtin> looks like the guy installing the GPU in masterblaster is ready, so I'll go ahead and take it offline in the next few minutes I think

20:58 < rcurtin> ah, he says actually that he'll do it in ~2-3 hours, so I'll wait to bring it down until then

20:58 < zoq> nice, hopefully it comes back in one piece

20:59 < rcurtin> yes, we'll see... :)

21:49 < Guest5588> how can I contribute as a beginner?

21:54 < zoq> Guest5588: Hello there! Have you seen http://www.mlpack.org/involved.html or http://www.mlpack.org/gsoc.html?

21:54 < zoq> Guest5588: We are always open for new contributions, so feel free to discuss/ask anything here, over the mailing list or via Github

21:54 < Guest5588> thanks

23:23 < rcurtin> ok, taking down masterblaster now... sorry for interrupted jobs...

23:30 < kris__> Okay i tried a couple of parameters for training with digits dataset.... https://usercontent.irccloud-cdn.com/file/C0YAul5c/Screen%20Shot%202017-08-23%20at%204.26.06%20AM.png

23:30 < kris__> The best i got was this....one but it is supposed to look like 7s and not 9's so.... i am not sure.

23:31 < kris__> https://usercontent.irccloud-cdn.com/file/5A4ebV67/Screen%20Shot%202017-08-23%20at%204.39.50%20AM.png

23:33 < kris__> I am not able to run mnist dataset even 4000 images we extracted on my machine with given set of pramas in enough time to do the parameter tuning for that.