00:35 < wenhao> lozhnikov: Hi Mikhail! About the work product submission for GSoC, how about I write it in the form of a blog and send you the link to preview before submission?

10:47 < zoq> wenhao: If you like you can write the report in the form of a blog post, examples from previous years:

10:47 < zoq> - http://www.mlpack.org/gsocblog/implementation-of-tree-types-summary.html

10:47 < zoq> - http://www.mlpack.org/gsocblog/deep-reinforcement-learning-methods-summary.html

10:47 < zoq> - http://www.mlpack.org/gsocblog/summary-of-lsh-changes-for-gsoc-2016.html

11:14 < zoq> rcurtin: Any ideas regarding the failing LMNNAccuracyTest test case? Not sure should we disable the test for now?

11:38 < akhandait> zoq: Sorry for asking about this again, but I am still a little confused about why we do this:

11:38 < akhandait> https://www.irccloud.com/pastebin/iRXk7A8d/

11:39 < akhandait> sorry, that's wrong, this:

11:39 < akhandait> https://www.irccloud.com/pastebin/c7fKqTBV/

11:41 < akhandait> how is mappedError(which is from the next layer) related to padW and padH of the current layer. We do take padding into consideration here:

11:41 < akhandait> https://www.irccloud.com/pastebin/mEqe5r3r/

12:35 < zoq> Just checked the code, and there is no need to cut the error if padding > 1, the DGAN test should fail, and I think this is a result of a wrong kernel size.

13:07 < akhandait> zoq: To be clear, did you mean that the DCGAN test passes because of a wrong kernel size or did you mean that the current code(dividing by paddding) was a result of wrong kernel size?

13:11 < zoq> akhandait: No I guess, it slipped through the last modifications.

13:12 < akhandait> zoq: Okay, I tried removing /padH and /padW, but the DCGAN test still passes and so does the convolutional network test(which I modified to non zero padding)

13:13 < akhandait> How are they both passing either way?

13:14 < zoq> akhandait: Did you build with DEBUG=ON, the test failed on my side.

13:14 < akhandait> zoq: Oh, I think I built with the default configuration

13:15 < zoq> in this case DEBUG is set to OFF

13:15 < zoq> "If you run CMake with no options, it will configure the project to build with no debugging symbols and no profiling information"

13:16 < akhandait> I will try it again with debug on

13:25 < akhandait> zoq: Yes, it failed now. So, with DEBUG off, what source code are the tests compiled with?

13:29 < zoq> With DBEUG=OFF the optional bounds check (armadillo) is disabled at compile-time to get more speed.

13:30 < akhandait> zoq: Oh, now I understood.

13:31 < akhandait> zoq: Thanks for the reviewing the PR.

14:27 < akhandait> zoq: I think there has been a confusion here. The DCGAN test fails when we DON'T divide by padding. So, have you been meaning to say that we don't need to divide by padding or were you saying we DO need to divide by padding? Just to inform, the ConvolutionalNetworkTest(I modified the first layer to have padding 2) passes either ways(if we divide by padding or not).

14:33 < zoq> akhandait: Haven't looked into the DCGAN test issue, my first guess is it uses a wrong kernel/padding/stride parameter, which causes the issue, there should be no need to devide by the padding value.

14:34 < zoq> ShikharJ: Do you see any reason to do so?

14:35 < akhandait> zoq: Okay, got it.

14:36 < ShikharJ_> zoq: Sorry, I didn't follow the context of the conversation. akhandait can you send me a link to the changes that you're making, and maybe then I can try to provide an explanation?

14:39 < akhandait> ShikharJ: I am not making any changes here, I couldn't understand why we have divided by padding here:

14:39 < akhandait> https://github.com/mlpack/mlpack/blob/b52f4b87d3a30cd345137c36edaa26aa48b2fbdd/src/mlpack/methods/ann/layer/convolution_impl.hpp#L245

14:39 < ShikharJ_> zoq: Also, please let me know if any changes are being made to the convolutional infrastructure of mlpack. I'd like to review the same before we merge anything new. I stopped following the repository, because my e-mail was flooded with conversations.

14:43 < zoq> ShikharJ_: For sure, you are the expert, so any input is much appreciated.

15:03 < ShikharJ_> akhandait: I think it was done to reduce the error map, in order to leave out the area of the error map which was convolved out of additional padding. But I do think there is an issue here now that I think of it. Let me check that again.

15:17 < ShikharJ_> akhandait: I think this was only done keeping the speedup benefits. Ideally, there should be no reduction whatsoever, even if it gives a speedup, because we're deliberately losing a bit of information and saturating ourselves only to the input cells when it comes to generating the gradients.

15:20 < akhandait> ShikharJ_: That's what I was thinking, we are losing gradients due to this. I will remove it in my next PR.

15:21 < akhandait> ShikharJ_: As Marcus said, do check the DCGAN test. It fails when we remove the division by padding.

15:22 < ShikharJ_> akhandait: That is expected, because we'll also remove the reduction that is happening later in the routine. But I'm also sensing a problem with removeing the reduction.

15:22 < ShikharJ_> *we'll have to

15:23 < ShikharJ_> akhandait: Let me make the changes and add a bit of comments.

15:24 < akhandait> ShikharJ_: Sure, I will just make the changes locally till your PR gets merged and work with it.

15:26 < ShikharJ_> akhandait: I'll have to reduce the error mapping in a way that the matrix multiplication doesn't produce zeros (imagine a 1x1 input with padding 2 and kernel 2x2, the first few convolutions would only produce zeros on the output matrix, and no matter how you update the gradients, the error will stay the same because it would always lead to zeros).

15:28 < ShikharJ_> akhandait: I think it was because of this philosophy that this reduction had been implemented, and I think zoq had also mentioned this once to me.

15:34 < ShikharJ_> akhandait: The padded zeros are hardly ever useful for generating the gradients.

19:14 < zoq> ShikharJ_: Thanks for looking into the Gradient method.