06:42 < ShikharJ_> zoq: I'm just curious, have you ever come across a paper which made use of non-square kernel sizes?

07:32 < akhandait> zoq: ShikharJ: I think there is one more thing we need to take care of. In the Gradient() function we don't deal with the case when size of gradientTemp > size of output. That's the reason DCGAN test is failing. I think we need to check if size of gradientTemp > size of output before convolving and pad the inpulSlice with (size of gradientTemp - size of output) number of zeros.

07:33 < akhandait> I did that locally and the test passes now

11:42 < zoq> ShikharJ_: "Learning on Graphs in the Game of Go" does use a non-square kernel, but that's pretty unusual. If you knew that all your features would have a rectangle shape, then it might be a good idea to go with a rectangular kernel, however, in most cases, you don't know the feature shape. It would be interesting to explore a circular kernel.

11:43 < zoq> akhandait: I see, that is basically what has been done with the padding parameter.

13:15 < TheKnight> Can a beginner also contribute.

13:15 < TheKnight> If so, how and where to start?

13:23 < TheKnight> Hi

13:23 < TheKnight> I'm E. Aakash, 2nd year undergraduate student of IIT Bombay.

13:23 < TheKnight> I am interested in C++ programming and wanted to know how machine learning algorithms are implemented in it.

13:53 < zoq> TheKnight: Hello, everyone is more than welcome to contribute; http://www.mlpack.org/involved.html should be helpful.

13:57 < zoq> ShikharJ: Looks like the RBM test sometimes failed: http://ci.mlpack.org/view/mlpack/job/mlpack%20-%20git%20commit%20test/lastCompletedBuild/testReport/

15:05 < Shikhar> zoq: I'm inclined to think these errors are random, one was an integer divide by zero error, and the other was a memory exception. I don't think these have to do with model performance.

15:05 < Shikhar> zoq: Let's wait and see if these persist, maybe then we can ascertain what's causing them?

15:08 < zoq> Shikhar: We should definitely look into the issues: http://ci.mlpack.org/view/mlpack/job/mlpack%20-%20git%20commit%20test/ they almost failed each time, the commit job runs.

15:09 < zoq> The divide by zero issue is probably easy to fix.

15:15 < Shikhar> zoq: The RBM PR was merged over two weeks ago, and these errors came up about a week ago, so maybe some changes on the internally used functionality caused this.

15:17 < Shikhar> zoq: Ah, it seems that these tests were failing from the start.

15:17 < zoq> Shikhar: Right

15:17 < Shikhar> That's a bit weird considering that the tests were in fact passing.

15:18 < Shikhar> zoq: Can you tell me the cmake configuration of the test bench, so that maybe I can reproduce the error?

15:21 < zoq> The build does use the default settings, so just 'cmake ..'

15:23 < zoq> Currently I run:

15:23 < zoq> while(true); do bin/mlpack_test -t RBMNetworkTest/BinaryRBMClassificationTest; sleep 1; done

15:24 < zoq> to see if I can reproduce the issue.

15:25 < Shikhar> zoq: Ok, lets' seeif I'm able to reproduce the error .

15:25 < zoq> Each run passed so far.

15:26 < zoq> Perhaps we should see if we could reproduce the issue if we build with gcc instead of clang.

15:27 < Shikhar> zoq: How is git-commit test different from docker mlpack nightly build?

15:28 < zoq> Shikhar: The test failed for the nighlty build as well, looks like the build only fails if the number of tests is > 2.

15:29 < zoq> But the nightly build tests different boost/armadillo versions.

15:31 < Shikhar> zoq: I checked through the code for BinaryRBM and its associated test. We only use the divide operation twice, that too on initialized integers, so my suspicion is that the error is occuring some place internally with the data we are passing.

15:33 < zoq> Shikhar: Agreed, the test code looks fine.

19:25 < zoq> Shikhar: Wasn't able to reproduce the issue, any luck on your side? Also, do not forget to use another random seed for each run(math::RandomSeed(std::time(NULL));).