07:35 < luffy1996> zoq: I think bandit algorithms should be implemented in mlpack. Bandit algorithms gives a real taste of reinforcement learning to the first timers. It will be great for the crowd using mlpack for machine learning. More over going through scikit-learn I saw that reinforcement learning is not supported by them. It will be great if mlpack does so, because it will easily benefit the first timers in the field. Hence I

07:35 < luffy1996> believe it should be added to mlpack. I think I will complete the implementation of bandit algorithms by March End. Presently I am busy with my semester exam. I will be free in a week after which I will implement a simple multiarm bandit algorithm in mlpack. What are your ideas ?

07:37 < luffy1996> More over multiarm bandits can easily be run on a CPU. So I think we should implement this :)

09:43 < dk97[m]> Hi there! Dakshit again.

09:44 < dk97[m]> zoq: rcurtin I see that adamax and Nadam are not implemented in mlpack. Is it alright if I implement these optimizers?

09:44 < dk97[m]> https://arxiv.org/pdf/1412.6980v8.pdf See section 7, it mentions adamax.

09:45 < dk97[m]> http://cs229.stanford.edu/proj2015/054_report.pdf <- nadam

09:45 < dk97[m]> It will help in expanding the variety of optimizers available to the mlpack community.

12:11 < travis-ci> yashsharan/models#3 (master - 260811d : yoloman): The build has errored.

12:11 < travis-ci> Change view : https://github.com/yashsharan/models/compare/24e2ad5d7f96...260811d54405

12:11 < travis-ci> Build details : https://travis-ci.org/yashsharan/models/builds/345904635

12:25 < travis-ci> yashsharan/models#4 (master - f9fa9ae : yoloman): The build has errored.

12:25 < travis-ci> Change view : https://github.com/yashsharan/models/compare/260811d54405...f9fa9aebc8e6

12:25 < travis-ci> Build details : https://travis-ci.org/yashsharan/models/builds/345906001

12:41 < zoq> dk97[m]: Adamax and Nadam is already implemented: http://www.mlpack.org/docs/mlpack-git/doxygen/namespacemlpack_1_1optimization.html

12:43 < caladrius[m]> zoq: Hi! Is there any functionality in mlpack which lets the user have independent biases for different strides of convolution filters?

12:49 < zoq> caladrius[m]: Currently, there is just the standard convolution operator implemented, I can't see a simple method to e.g. add it with a single line of code. So we would have to go and e.g. implement it as another layer or integrate it into the convolution class.

12:49 < zoq> luffy1996: Do you have something specific a specific Bandit algorithm in mind? I think the selected algorithm should fill some niche, e.g. faster as RL or more accurate. Best of luck with your exams.

12:54 < caladrius[m]> zoq: There's a FlexibleReLU activation function which tackles this problem to some extent. Here's the paper for this: https://arxiv.org/pdf/1706.08098.pdf . It adds a bias to the rectifier function and improves the performance of convolution layers.

12:55 < zoq> caladrius[m]: Oh nice, that would be easy to add.

12:56 < caladrius[m]> Yeah. Could I try it out then?

12:57 < zoq> Sure, please feel free.

12:58 < caladrius[m]> Thanks!

13:00 < pc14> Hi people, I am a prospective student candidate for GSoC. I was wondering if someone could point me to possible warm-up tasks. Thanks (https://drive.google.com/file/d/18oh0ssYriN3k9dWR3NAyuqq993EgoO4D/view)

13:05 < dk97[m]> thanks zoq sorry to miss that

13:07 < dk97[m]> could you have a look at the PR I submitted? If the implementation is alright, I can write a few tests for the same.

13:12 < zoq> dk97[m]: I'll take a look at the PR once I get a chance, but please be patient there are a bunch of PR's open, so this might take some time, but we will respond as fastest as possible.

13:13 < dk97[m]> sure no problem! :)

15:12 < manthan> i subscribed o mlpack list but i didnt get the confirmation mail. Approximate how many days will it take for the same? Also, is there any other medium to introduce ourselves to the community?

17:16 < travis-ci> yashsharan/models#6 (master - b715877 : yoloman): The build has errored.

17:16 < travis-ci> Change view : https://github.com/yashsharan/models/compare/bb1eeced29e6...b71587735e1f

17:16 < travis-ci> Build details : https://travis-ci.org/yashsharan/models/builds/345969820

17:31 < moksh> Hey @zoq, I am quite new to reinforcement learning, so as you suggested, I was thinking of implementing the SARSA algorithm, since it is quite similar to Q-Learning, will that be fine?

17:38 < travis-ci> yashsharan/models#6 (master - b715877 : yoloman): The build has errored.

17:38 < travis-ci> Change view : https://github.com/yashsharan/models/compare/bb1eeced29e6...b71587735e1f

17:38 < travis-ci> Build details : https://travis-ci.org/yashsharan/models/builds/345969820

18:06 < dk97[m]> zoq: there is another layer proposed in the Self Neural Networks paper- alpha-dropout- a variant of dropout that they used. Is it alright if I implement that as a follow-up to my PR? Then an implementation of an SNN could be done.

18:06 < dk97[m]> rcurtin:

18:07 < luffy1996> zoq: When it comes to implementation of bandit, I feel it will be faster as it will be coded on C++. The benefit which I see from my side is, it will make life easier for the crowd trying to implement these algorithms using reinforcement learning. Apart from this I see that various algorithms like Sarsa or TD(lambda) {eligibility traces} , monte carlo methods are yet to be implemented in mlpack. Coming from

18:07 < luffy1996> multiagent reinforcement learning background I feel that the implementation for IQL ,DDPG should also be present. The number of games defined under RL should increase to support diversity in playing games. One should also not forget the classic bellman equations to solve MDPs. There are a lot in RL literature which can be implemented on mlpack. The questions stands how we should prioritise this?

18:07 < luffy1996> Having said that I do understand this requires time and effort. It will be great to know your inputs and how you feel about implementation of any of these . My current plans for March includes implementation of bandit algorithms using RL and adding one or two game in the kitty . This will necessarily help the user to implement algorithms in diverse environment. Thanks for wishing me luck. Hope the semesters go well

18:07 < luffy1996> :)

18:13 < rcurtin> luffy1996: my take on implementing bandits as a project is that this isn't traditionally functionality that mlpack has supported

18:13 < rcurtin> it would be okay to add bandit algorithms

18:13 < rcurtin> but the key to any proposal would be that at the end of the summer we have a fully finished product that users can easily use

18:13 < rcurtin> this has been done successfully in the past, with e.g. the CF code

18:14 < rcurtin> but I would say that it is very important for us to provide complete functionality, so I might suggest that you spend your time planning your proposal carefully

18:14 < rcurtin> as opposed to quickly implementing bandit algorithms to be merged into mlpack

18:15 < rcurtin> dk97[m]: I don't have a problem with that so long as the alpha-dropout layer is useful outside the context of SNNs

18:28 < dk97[m]> alpha-dropout was more to be used with SELU activation function because of a different default low variance value

18:28 < dk97[m]> rcurtin:

18:29 < dk97[m]> Although, since it is a variant of dropout, its usage outside of SNN should not take a hit.

18:31 < navabhi> Hi everyone !! I cloned the latest mlpack github repo and compiled it on my system in debug mode.

18:31 < navabhi> I was just going through the tutorials for different log levels in mlpack from http://www.mlpack.org/docs/mlpack-git/doxygen.php?doc=iodoc.html .

18:32 < navabhi> but on running the example, I did not see any Debug or Info message on my terminal.

18:34 < navabhi> I compiled mlpack with -D DEBUG=ON; and the above file as "g++ example.cpp -std=c++11 -lmlpack -g -rdynamic --verbose". Am I missing anything here ?

19:44 < daivik> navabhi: after building, did you install mlpack by running make install?

19:45 < navabhi> yes, I did that too.

19:47 < daivik> Strange, there were no errors during the build? are you sure you configured using -DDEBUG=ON and then ran both make and make install?

20:54 < navabhi> Yes, there were no errors. A couple of warnings though.

20:55 < navabhi> The CMake output on running 'cmake -DDEBUG=ON ../' can read here: https://pastebin.com/jcHq5Git

21:01 < daivik> you are running the program with the --verbose flag?

21:03 < daivik> also, you should have got an error on the CLI::ParseCommandLine(argc, argv) line -- what did you do about that?

21:08 < navabhi> just commented it out; also tried --verbose flag but got the same output

21:10 < daivik> https://thepasteb.in/p/oYhlGqpjjpBHZ -- > try running this

21:22 < navabhi> I tried running your script, but still got the same output. (I also had to link boost to run the code)

21:28 < daivik> yes, did you run with the --verbose flag?

21:28 < daivik> you should atleast get the INFO outputs

21:39 < navabhi> yes, I used this command: 'g++ test_log.cpp -std=c++11 -lmlpack -lboost_program_options --verbose'

21:40 < navabhi> I had to pass the verbose flag at runtime. Sorry for the mistake. I am now getting the INFO logs

21:49 < navabhi> The log of the make command can be accessed from https://pastebin.com/ybTHw4y4, if it helps

21:52 < daivik> I'm also not getting any debug outputs -- which is strange because I get debug outputs when I run tests.

21:56 < daivik> also, when I try something like mlpack::Log::Debug.ignoreInput=true; it tells me that Debug is a NullOutStream .. which means that DEBUG is not set while compiling. This is probably an issue. rcurtin, zoq - could you help us out please?

21:56 < navabhi> well, I don't get debug outputs when running tests too which is why I was trying this example program. I do get info outputs on setting log_level as all.

21:58 < daivik> I compiled with -DDEBUG=ON and -DTEST_VERBOSE=ON and I see DEBUG outputs for tests.. maybe you can try -DTEST_VERBOSE=ON and that will help

22:04 < daivik> sorry, but I have to go now... I'll look into this more tomorrow. Will let you know if I find something :0

22:04 < daivik> * :)