00:18 < rcurtin> zoq: !!!!!

00:18 < rcurtin> I would say that is more than somewhat satisfying :)

00:18 < rcurtin> but maybe I just get really excited by progress bars

00:19 < rcurtin> also, that's an interesting name for an Armadillo release, "Fomo Spiral"

00:19 < rcurtin> I assume it refers to 'fear of missing out' :)

03:00 < Yashwants19> Hi rcurtin: should we also update readme section on github.

03:00 < Yashwants19> ?

03:04 < rcurtin> Yashwants19: yeah, in the process of that now

03:04 < rcurtin> not everything is perfectly automated quite yet, but I am getting closer with each release :)

03:20 < travis-ci> mlpack/mlpack#7069 (mlpack-3.1.1 - 78cab8b : Ryan Curtin): The build passed.

03:20 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/603a11f29a16^...78cab8b4a21b

03:20 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/537595489

04:28 < favre49> rcurtin: The blog page mlpack.org/blog seems to be a bit broken, the website template isnt loading i think, just the text.

05:23 < travis-ci> mlpack/mlpack#7072 (master - 584f853 : Ryan Curtin): The build was broken.

05:23 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/2414da2c43fa...584f8533f20a

05:23 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/537604184

07:47 < jenkins-mlpack2> Project docker mlpack nightly build build #338: NOW UNSTABLE in 3 hr 33 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/338/

09:17 < zoq> rcurtin: Conrad has all the good names :)

09:42 < mulx10> zoq: I am facing issues with integration of load image (https://pastebin.com/bDkTycYC). produces error (https://pastebin.com/MEXG2m4S)

09:43 < mulx10> the implementation are (https://github.com/MuLx10/mlpack/blob/LoadImage/src/mlpack/core/data/load_img_impl.hpp) , (https://github.com/MuLx10/mlpack/blob/LoadImage/src/mlpack/core/data/load_img.hpp)

09:44 < mulx10> also the updated cmake list (https://github.com/MuLx10/mlpack/blob/LoadImage/src/mlpack/core/data/CMakeLists.txt).

09:45 < mulx10> Am I missing some place where I need to add the new files?

09:50 < zoq> mulx10: Is stb_image.h in your include search path e.g. /usr/local/include/?

09:51 < mulx10> zoq: Yes, I have added it via (https://github.com/MuLx10/mlpack/blob/bd7ad7d5778933f573b646600680d178ec54fd29/CMakeLists.txt#L329)

09:53 < zoq> mulx10: I guess, you did run 'make install' as well?

09:53 < zoq> mulx10: Can you search for the file?

09:53 < mulx10> Yes, I did starting from cmake to make install

09:54 < mulx10> zoq: stb_image.h file?

09:55 < zoq> right

09:55 < mulx10> locate stb_image.h gives me this

09:56 < mulx10_> . /home/nirala/mlpack/build/include/stb_image.h

09:56 < mulx10_> /home/nirala/mlpack/build/include/mlpack/core/data/stb_image.h /home/nirala/mlpack/src/mlpack/core/data/stb_image.h /usr/include/mlpack/core/data/stb_image.h

09:57 < zoq> okay, looks good

09:57 < zoq> let me clone the repo

09:58 < mulx10_> Here's the link https://github.com/MuLx10/mlpack/tree/LoadImage

09:59 < zoq> also, do you get the same error if you put the code into a boost test case e.g. https://github.com/mlpack/mlpack/blob/master/src/mlpack/tests/load_save_test.cpp?

10:01 < mulx10_> I guess there is some linking error

10:01 < zoq> yeah, so if it works in the test case, that might be the issue.

10:02 < mulx10_> I haven't tested in boost, I am doing it now.

10:04 < zoq> oh wait, there is a missing: // Include implementation.

10:04 < zoq> #include "load_img_impl.hpp"

10:04 < zoq> at the end of load_img.hpp

10:05 < zoq> Here is an example: https://github.com/mlpack/mlpack/blob/584f8533f20a4b94831d6b68ebb100bcbee573b7/src/mlpack/core/data/dataset_mapper.hpp#L201

10:06 < mulx10_> Oh okay.

10:08 < mulx10_> Thanks :) If it works I'll let you know.

10:18 < mulx10_> zoq: It worked thanks!

10:20 < zoq> great :)

12:27 < rcurtin> happy first day of GSoC :)

12:58 < KimSangYeon-DGU> Wow :)

12:59 < KimSangYeon-DGU> sumedhghaisas_: Hey Ghaisas, Please let me know, if you are ready :)

13:08 < KimSangYeon-DGU> I've completed the plotting of the probability surface of QGMM. I can't find any difference between QGMM and GMM. So I guess we can check the difference when training the models.

13:09 < KimSangYeon-DGU> You can check it at https://github.com/KimSangYeon-DGU/GSoC-2019/tree/master/Research/Probability_Space

13:18 < KimSangYeon-DGU> So, I think it's a good start :)

13:19 < KimSangYeon-DGU> because we found the two probabilities are similar in initial stage.

15:53 < ShikharJ_> sakshamB: Toshal: Let's make this quick, I have a flight to catch. So if you guys are here, let me know and we can start to discuss?

15:53 < sakshamB> yes I am here

15:55 < ShikharJ_> Okay, so what do you have in mind regarding mini-batch discrimination?

15:56 < sakshamB> I looked at existing implementations of minibatch discrimination in python

15:57 < sakshamB> It is implemented as a layer and I think we just add it as a layer in mlpack

15:58 < sakshamB> I am working on finding the derivative for backpropogation..

15:58 < ShikharJ_> Okay, I think it would be pretty straight-forward with the implementation once you have found the gradients.

15:59 < ShikharJ_> I recently reviewed you PR on highway networks, and I think it is largely complete. So that should be done this week as well.

16:00 < ShikharJ_> sakshamB: Have you thought about how you're going to test out the layer?

16:00 < sakshamB> yes. just wanted to ask if you knew any resources to compute the formula for backpropogation or do i find it manually and verify?

16:00 < ShikharJ_> sakshamB: Manual computation is what we have predominantly relied on if in case the paper for the technique does not mention it explicitly.

16:01 < ShikharJ_> We can discuss this in detail on the PR itself. I'm not sure if we can write LaTeX code in the IRC.

16:02 < sakshamB> hmm..numerical gradient is trivial. but more importantly we need to check and verify for the quality of images maybe through inception score and other metrics.

16:02 < sakshamB> numerical gradient test*

16:03 < ShikharJ_> sakshamB: That would be a task that you can take up later and develop this feature further when you have implemented Inception Score metric.

16:05 < sakshamB> alright so, would we be able to merge the layer without verifying the quality of the images?

16:06 < ShikharJ_> For now, let's focus on getting a basic layer template code up as a PR, and we'll discuss there. I have a couple of ideas regarding the testing of this layer, but I'll need broader discussion to gain the confidence that they're viable.

16:06 < ShikharJ_> sakshamB: I'm not sure what you mean by "without verifying the quality of images".

16:07 < sakshamB> computing the inception score to compare the performance

16:07 < ShikharJ_> Ah, sorry, I didn't keep in mind your previous comment :)

16:09 < sakshamB> alright so I can start working on the implementation. Will try to open a PR soon.

16:09 < sakshamB> when is our next meet again?

16:09 < ShikharJ_> Yeah, if Inception score is essential, I think we can wait till that is implemented. Let's focus on these tasks with a deadline of two weeks, and you can take on the rest of your choice of work later.

16:10 < ShikharJ_> I'm assuming you're interested in taking Inception Scoring as the next piece of work. Let me know if it isn't and maybe we can work that out as well.

16:11 < sakshamB> yes, I don’t mind working on inception scoring next since a lot of work for GANs relies on it

16:12 < ShikharJ_> sakshamB: Great. We decided on bi-weekly meets right? Monday and Friday at 9.30 pm? Let me know if you want a change there as well?

16:13 < sakshamB> no, that is perfectly fine with me. i just wanted to confirm. thanks anyways. :)

16:13 < ShikharJ_> I'll be available on IRC throughout as well. I just messed up my bouncer configuration, so I'm checking the logs. Hopefully I can get it working soon.

16:14 < ShikharJ_> Okay, anything else you wanted to discuss? I think Toshal would be missing this meet, so I'll hang around for a while and then I'll be off :)

16:15 < ShikharJ_> Also, if you feel there is some other work you wish to take up now, feel free to push the code and we'll review.

16:17 < sakshamB> no, I will ask if I have any doubts later on. thanks again

16:18 < ShikharJ_> Great :)

16:23 < aleixrocks[m]> Hi all! Is there any option to use mlpack/ensmallen/armadillo with out-of-core support? I have read several mlpack tutorials and examples but apparently data is always read at the beginning of the execution and written at the end of it. I would like to work with datasets bigger than my available system's memory. Well, in fact, what I'm truly interested is in studying the storage device impact on ML/DL executions (mostly

16:23 < aleixrocks[m]> the SGD part) for my PhD.

16:27 < rcurtin> aleixrocks[m]: it's easy with mmap() but I'm about to get lunch so I can answer better when I am back

16:28 < rcurtin> there is a paper... I forget its name... by Andy Fang and Polo Chau at SIGMOD 2016 I think, called 'M3'? (that's only part of the name)

16:28 < rcurtin> you could look at papers citing the JMLR mlpack paper to find it

16:29 < rcurtin> anyway that might be helpful, I'll explain more about how to do it when I get back from lunch

16:30 < ShikharJ_> aleixrocks[m]: This might be of interest to you: http://harsha-simhadri.org/pubs/nsdi19_final.pdf

16:33 < aleixrocks[m]> rcurtin: Thank you very much for your reply! I'm familiar with the mmap syscall (I'm somewhat of a Linux kernel developer), but I will definitely appreciate some guidance on how to use it given that data management seem to be done through armadillo. I will check for this paper as well! Bon appetite!

16:42 < aleixrocks[m]> got it just as you said! M3: Scaling Up Machine Learning via Memory Mapping

16:43 < aleixrocks[m]> ShikharJ_ paper is also quite interesting, thank you so much!

16:45 < favre49> Is there any way i could initialize a struct or enum whose definition is a member of a template class?

16:47 < favre49> I'm trying to create a class that can wrap the RL environments(passed as a template param), and i need to initialize the Action.

16:47 < favre49> If i passed the template class as EnvironmentType, would EnvironmentType::Action action work?

17:22 < rcurtin> aleixrocks[m]: so, the trick is to wrap Armadillo around an mmap() pointer

17:22 < rcurtin> basically, what you want to do is create a file on disk that stores the packed double representation of the matrix

17:23 < rcurtin> i.e., each element stored as an 8-byte double, sequentially, in column-major order

17:23 < rcurtin> then, you can use mmap() to get a pointer to this file, and then use the Armadillo advanced constructor like this:

17:23 < rcurtin> double* mmapPointer = mmap(...)

17:24 < rcurtin> arma::mat mmapMatrix(mmapPointer, n_rows, n_cols, false /* don't copy memory */, true /* cannot change size */);

17:24 < rcurtin> now, note that you'll have to be really careful to make sure nothing tries to copy the mmapMatrix

17:24 < rcurtin> but otherwise, that should work without problems

17:38 < aleixrocks[m]> rcurtin: I see, this should definitely do the trick for my purposes. Thank you very much for your help and rapid response!

18:39 < zoq> favre49: What about passing an instantiation of the env as well? Something like: https://gist.github.com/zoq/52274326df17f7b18d4b4a1be16aea0e?

19:08 < rcurtin> favre49: I fixed the blog page, looked like it was loading the wrong image

19:09 < rcurtin> it'd be nice to transition that whole style to the new webpage style, but it's a lot of effort and if German is already redoing the website, maybe better to wait :)

19:09 < rcurtin> I think it's fine as-is anyway

19:09 < rcurtin> let me know if I missed something or there are other issues

19:16 < zoq> Okay, invited everyone to write to the blog repo as well.

20:52 < sreenik> zoq: Thanks for the invite