17:32 < rishabhgupta05> help

19:24 < qwertea> hello everyone :) I need to access to the mappings in a datasetinfo to display a text representation of a decision tree; I didn't find any method I can call to access the mappings, any idea?

19:25 < qwertea> I was adviced before to serialize it but I now need the mappings in my code

19:25 < qwertea> thanks in advance!

19:33 < zoq> qwertea: If you know the value you could use: http://www.mlpack.org/docs/mlpack-2.2.4/doxygen/classmlpack_1_1data_1_1DatasetMapper.html#a18ecaf8879808727e9165dc26b069564

19:33 < zoq> If not we could add a simple function that returns all mapping, maybe rcurtin can say more about this.

19:35 < qwertea> unfortunately I don't know the value :/ a function to get all mappings would be great!

19:38 < zoq> Do you mind to open an issue on github for this? Maybe someone likes to work on this; I could also implement the addition.

19:41 < qwertea> sure no worries! it also looks relatively simple so I could also implement it

19:41 < zoq> qwertea: Great!

20:25 < sumedhghaisas_> zoq: Hey Marcus, I am doing a clean build right now.

20:26 < sumedhghaisas_> while I do that, had couple of doubts

20:26 < sumedhghaisas_> I merged the copy task code in NTM to perform copy task. Even though there were not many conflicts

20:27 < sumedhghaisas_> I noticed that there is some extra code about 'rho' parameter... how do you want to fix this issue?

20:28 < sumedhghaisas_> also how much is the error in GradientNTMTest??

20:28 < zoq> We can remove that part, since you already fixed that in the GRU PR.

20:29 < sumedhghaisas_> zoq: shiiiiiiiiiit.... I got the problem

20:30 < sumedhghaisas_> it was in my initialization of controller in ann_layer_test

20:31 < sumedhghaisas_> I initialised the first layer of the controller wrong...

20:31 < sumedhghaisas_> should be '10 + 6'... not '10 + 5'... I was using 5 block memory before

20:31 < sumedhghaisas_> but then for testing I also tried 6...

20:31 < zoq> ah, I see, easy to fix

20:32 < sumedhghaisas_> I will commit now ... lets see if the online gradients pass

20:32 < sumedhghaisas_> wait... sorry it should be '10 + 10'...

20:34 < sumedhghaisas_> I think those network checks that we discussed that day, are necessary :P

20:35 < zoq> Agreed :)

20:36 < sumedhghaisas_> also could you help in setting up the NTM tests for copy task on the server? my computer becomes too slow with those tests running

20:38 < sumedhghaisas_> another issue... I installed the merged NTM and benchmark tasks... there was some error while building models repo. Something related to parallel_sgd... some implementation not being there. So I deleted the extra headers in sgd_impl.hpp and it worked.

20:38 < sumedhghaisas_> don't know if thats the right fix...

20:38 < sumedhghaisas_> but maybe those headers are added by mistake?

20:39 < zoq> The error should be fixed once https://github.com/mlpack/mlpack/pull/1077 is merged.

20:41 < sumedhghaisas_> ahh okay then

20:41 < zoq> About the copy task, do you want to use the code from the models repo or do you like to write a boost test case?

20:42 < sumedhghaisas_> I like the idea of 'models' repo... I would like to use that.

20:42 < sumedhghaisas_> should I send a PR there?

20:43 < sumedhghaisas_> I also changed some parameters in LSTM baseline... it is performing better now.

20:43 < sumedhghaisas_> I hope thats okay...

20:43 < sumedhghaisas_> I think the learning rate was quite high... cause the learning was very unstable.

20:44 < zoq> I think it would be neat if we could provide some pretrained models, and also show some nice examples how to use the model and the place for that would be the models repo.

20:44 < sumedhghaisas_> Also changed the optimizer to RMSProp...

20:44 < zoq> Ah nice :)

20:45 < sumedhghaisas_> So what do you propose?

20:45 < zoq> About the test?

20:46 < sumedhghaisas_> yeah. Add some simple copy task to boost test and major ones in models?

20:46 < sumedhghaisas_> ohh... you were saying both in models. Sorry I got the wrong idea...

20:48 < zoq> It would be great to have a unit test in mlpack for the NTM model (boost test case), we have to see if we can come up with something that runs reasonable fast on travis...

20:48 < sumedhghaisas_> ahh... yes I was thinking about that. You know we have recursive reber grammar test. Why not make it more recursive?

20:49 < sumedhghaisas_> So we define a recursive depth and get a reber grammar.

20:49 < sumedhghaisas_> it would be cool to check at which recursive depth the LSTM fails...

20:49 < sumedhghaisas_> that would also be the NTM model test

20:49 < zoq> Sounds interesting, nice idea.

20:51 < sumedhghaisas_> Okay I will do that first. Then lets setup the models thing. is that fine?

20:51 < zoq> Sure sounds fine, for me.

20:53 < sumedhghaisas_> okay. So did you get a look at the MemoryTest design? Do you think thats the correct design for testing memory gradients?

20:55 < zoq> Let me take a closer look at the MemoryTest tomorrow.

20:55 < zoq> About the models repo: Basically all you have to do is to copy the code from the ann/augmented/tasks folder. Clone the models repo adjust GenerateModel function and build with:

20:55 < zoq> cmake -DMLPACK_LIBRARY=/path/to/build/lib/libmlpack.dylib -DMLPACK_INCLUDE_DIR=/path/to/build/include/ && make

20:55 < zoq> and at the end run the run_copy_task.sh script.

20:57 < sumedhghaisas_> okay. Ahh yes... I think I have login to the server.

20:57 < sumedhghaisas_> but how do I select the machine on which to run the task?

20:59 < zoq> I think you should be able to ssh into masterblaster.mlpack.org?

21:00 < sumedhghaisas_> yup. I can.

21:02 < sumedhghaisas_> ohh... I just did lscpu there... its a single 72 core machine... I thought they are separate machine. Thats what we have here in the uni

21:02 < zoq> ah, yeah no it's a single machine

22:05 < ironstark> rcurtin: zoq: I will soon start working on benchmarking R so we will need to setup R.I know how to implement the algorithms in R (after the dataset is read). I might need help with the part where we send the command using python scripts to run the R script. I also need help with the changes we need to make to the Makefile for setting up and ensuring smooth functioning of R scripts. I also might need help with how to

22:05 < ironstark> return the runtime calculated back to the python script.

22:06 < ironstark> The first thing I need to know is how to setup R on the system. Once that is done I'll try to figure out other problems on my own by running the scripts, if I am not able to I'll take help

22:07 < ironstark> but the setup thing is something I am not sure how to achieve. Since R is a language and not a library so how do we set it up using make setup

22:22 < ironstark> I have also thought about benchmarking against some of the libraries like MachineLearning.jl( Julia), MLlib(Spark), Shark(C++) after I am done with R. Please let me know your thoughts on this.

22:24 < sumedhghaisas_> zoq: Okay the gradients are passing now. There is some compilation error in AppVeyor regarding arma::each_col function. is AppVeyor using older version of armadillo?

22:54 < zoq> sumedhghais: We build against 7.800.2, probably lack of lambda support ...

23:02 < zoq> ironstark: We could build R from source: https://cran.r-project.org/doc/manuals/r-release/R-admin.html#Building-from-source but we could also just install R via the package manager. I guess since we build everything from source we could at least test it out.