03:34 < Nie> Hello guys, I installed mlpack in my Ubuntu 16.04 with this command sudo apt-get install libmlpack-dev. But when I try using command like mlpack_fastmks, it says The program 'mlpack_fastmks' is currently not installed. You can install it by typing: sudo apt install mlpack-bin

03:35 < Nie> I'm confused about it.

05:43 < Prabhat-IIT> zoq: I've been trying to implement the SAGA optimizer that you've proposed to me with the Xavier initialization. Actually, I was not able to fully understand the code that authors have proposed but have understood the algorithm. I've implemented a straight forward version of it in python first. I want to know if the straight forward implementation can be implemented or not as it seems to be only 3 times slower than the actual one

05:44 < Prabhat-IIT> here's my short implementation in python https://pastebin.com/7ViQ86qW

05:45 < Prabhat-IIT> This implementation can be improved upon memory by using mini-batches for each epoch

05:49 < Prabhat-IIT> If it can be incorporated in mlpack like this version then I can go ahead and code otherwise I'll have to give it some more time for understanding authors implementation code. I think the difference will only be in supporting sparse gradients instead of dense ones!

13:17 < zoq> Prabhat-IIT: Hello, can you point me to the faster implementation? Also, if you like open a PR with the current implementation and we can figure out how to make it faster. For sparse support checkout http://www.mlpack.org/docs/mlpack-git/doxygen/optimizertutorial.html

13:19 < zoq> dk97[m]: Can you elaborate on the idea?

13:53 < namratamukhija> hi, I'm trying to run all test cases in mlpack using the command ./mlpack_test. I have the error - 55 failures detected in test suite "mlpackTest". Is this the right way of running the test suite? Most of the errors are of the type - fatal error in "DualTreeVsNaive2": Cannot load test dataset test_data_3_1000.csv!. Others are memory access violation and run_time errors.

13:55 < rcurtin> namratamukhija: be sure that you run the test from the build directory, not the build/bin directory

13:55 < rcurtin> you can also type 'make test' from your build directory

13:55 < rcurtin> you will see that test_data_3_1000.csv is in your build directory but not in the bin/ directory

14:02 < namratamukhija> rcurtin: thank you for pointing that out. I ran make test from the build directory - it seems stuck on Running tests... Start 1: mlpack_test from the past 2-3 mins. I'm not sure how long it's supposed to take. can you let me know?

14:08 < namratamukhija> rcurtin: It worked! Thanks! Took about 400 seconds. Not sure, but should there be some kind of verbosity option while running all tests? Something like .. running test file *name of test file*. If it is a good idea, I'll be happy to work on it.

14:09 < zoq> namratamukh: run with -p to get more information

14:16 < rcurtin> oops, sorry jenkins-mlpack... that is what happens when you assign two hosts to the same IP... :)

14:19 < namratamukhija> zoq: Thanks

14:58 < 17WAAVQQJ> mlpack/mlpack#4205 (master - 1572e91 : Ryan Curtin): The build has errored.

14:58 < 17WAAVQQJ> Change view : https://github.com/mlpack/mlpack/compare/9f69eb6cbfe8...1572e9147646

14:58 < 17WAAVQQJ> Build details : https://travis-ci.org/mlpack/mlpack/builds/348214118

15:19 < Prabhat-IIT> zoq: Here's a faster implementation given by the authors themselves in cython https://arxiv.org/pdf/1407.0202.pdf (checkout at the end) Its for sparse least squares and ridge regression

15:44 < rcurtin> ok, all of the build infrastructure is back online, including the sun systems

15:44 < rcurtin> I wish it hadn't taken so long, but now we should not need to have any more downtime for quite a while (...I hope...)

15:46 < dk97[m]> @zoc I was thinking of making a separate folder for loss functions, similar to how there are different folders for activation functions and initialization.

15:48 < dk97[m]> The loss folder would contain implementation of different loss functions.

15:49 < dk97[m]> Like the MSE, KL divergence, cross entropy

15:54 < dk97[m]> This could be useful in localising the loss functions to one place. Like right now KL Divergence loss is defined in the sparse auto-encoder implementation, but the same loss will be need in VAE as well.

16:02 < zoq> dk97[m]: I see, I'm fine with moving the existing ann layer into a separated folder.

16:03 < zoq> rcurtin: awesome

16:03 < zoq> Prabhat-IIT: Okay, thanks, will take a look at the paper.

16:09 < dk97[m]> zoq: Actually, I was thinking of having the loss function inside the ann folder.

16:10 < dk97[m]> ann->loss_functions->MSE.hpp

16:10 < dk97[m]> something like this

16:10 < zoq> dk97[m]: sounds good

16:10 < dk97[m]> So should I open an issue?

16:10 < dk97[m]> Or just make a PR?

16:14 < zoq> dk97[m]: If you like to work on this one, there is no need to open an issue first.

16:15 < dk97[m]> Okay then, tnanks!

17:38 < aman_> hello contributors! I went through goals and work done at mlpack and I believe my interests and skill overlap to a great extent with this orgs'. I really liked your project ideas, especially one about VAE and other about Reinforcement learning. I wanted to get my hands dirty to be ready to take up one of these projects during the summer as a gsoc project. Any heads up on how to start and go about it. Thanks in advance!

17:38 < aman_> Btw, I am Aman from IIIT-H, India pursuing CSE and have more than 3 years of experience with coding and have completed multiple ML, DL projects. :)

17:48 < zoq> aman_: Hello, www.mlpack.org/involved.html and mlpack.org/gsoc.html should be helpful.

17:52 < aman_> Thanks zoq , I have already built mlpack on my local machines. I will start off by solving some of the issues then.

17:52 < zoq> aman_: Sounds good.

17:53 < aman_> cool

18:25 < samidha1705> Hey, I've been trying to work on PSO algorithm and I've written down the psuedo-code of the PSO algorithm. What does the proposal 'exactly' want?

18:27 < zoq> samidha1705: Hello there, have you seen https://github.com/mlpack/mlpack/wiki/Google-Summer-of-Code-Application-Guide?

18:27 < samidha1705> Yes, I have!

18:27 < samidha1705> I specifically meant the technical details.

18:29 < samidha1705> Do I have to write the code and submit it? Can I send the technical proposal to any of the mentors right now?

18:32 < zoq> There is no need to write pseudocode, if you lay down the general idea this is just fine, the more important part is the interface, e.g. how it the project fits into the existing codebase.

18:32 < zoq> Once the student application period starts, mentors are able to take a look at the proposal and give feedback if the student enables the setting.

18:37 < samidha1705> Thanks but I think with optimizers as the category, the PSO algorithms will just be another addition to that category like other optimization algorithms.

18:37 < samidha1705> Do I need to provide how it will fit into the exisiting codebase from scratch or can I take the help of the pre-made optimizers algorithms?

18:39 < zoq> That is true for the standard PSO method, in that case you can just point that out. The constrained method might need some modifcations.

18:56 < samidha1705> That clears my doubt, thanks!

21:49 < rcurtin> zoq: seems like travis took even longer with ccache... does ccache give a speedup on your system? maybe it is random noise with travis

21:50 < rcurtin> I noticed PR #1289 built just fine within the time limits: https://travis-ci.org/mlpack/mlpack/builds/348160860?utm_source=github_status&utm_medium=notification

21:51 < zoq> I can see some speedup on my system but it looks like that's not the case on the travis machine.

21:52 < zoq> I'll test it one more time with clang, and close the PR afterwards.

23:46 < zoq> Besids the test failures, this are some pretty good timings: https://travis-ci.org/mlpack/mlpack/builds/348455878