02:29 < jeffin143_> lozhnikov : never knew what a specialization meant, went through some article and came up with this, Please take a look and let me know if this is what you were looking for

02:30 < jeffin143_> Here it is : https://pastebin.com/FF8KhmwA

02:30 < jeffin143_> Also, if this the what you were looking for, then how to enable it for specific boost version

02:32 < jeffin143_> Should i do like i did for boost::string_view, include boost_version.hpp and write a if condition (version < 16100) and fill in the stuff inside the if.

02:33 < jeffin143_> Also which file should have the specialization defination, since i guess string_view and map would be used in all of our string utility function

02:33 < jeffin143_> i guess the extra backported file which we just wrote that includes some file may be the write place !

02:35 < jeffin143_> right* :)

04:59 < lozhnikov> jeffin143_: Yes, I meant exactly that. I think you can use the same file boost_backport_string_view.hpp for the definition.

05:46 < jeffin143_> lozhnikov , rcurtin , zoq : could you please help me debug this code : https://pastebin.com/ssWF1nLb

05:47 < jeffin143_> I guess if i am correct the string_view hold the pointer to the string

05:47 < jeffin143_> and thus inserting v[1] or v[2] should work correctly but there is some error

05:49 < jeffin143_> but the output is , the map is unable to print the first element since it lost its address somehow : https://ibb.co/0jHKCrS

06:01 < lozhnikov> jeffin143: I guess it happened due to push_back().

06:03 < jeffin143> lozhnikov : how should I resolve it

06:03 < lozhnikov> jeffin143_: Try to allocate at the declaration. (std::vector<std::string> temp(3);). In that case it works.

06:04 < jeffin143> No , I can't because, I really don't know wether I have to insert the string or not based on a condition

06:04 < lozhnikov> i.e. push_back() copies the elements sometimes and the views become invalidated.

06:04 < jeffin143> I am doing , since you suggested to save the token in a vector and then insert it's string view

06:05 < lozhnikov> The easiest way to fix that is to consider std::list instead of std::vector

06:06 < jeffin143> Umm , so I would use STD::list to store the token I need to insert inside the map and then insert the view of list inside the map , that way if the input is cleared also we would have our map

06:06 < jeffin143> Thanks :)

06:06 < jeffin143> Btw how do you so quickly come with an answer , for my bug , I was trying hard to understand what the issue is

06:07 < lozhnikov> jeffin143: I think you can use std::deque as well. Maybe std::deque would work faster than std::list.

06:08 < jeffin143> I agree with you, let me see

06:10 < lozhnikov> jeffin143: I just tried to change m[temp[temp.size()-1]] by m[strview] and that worked. So I decided that the views become invalidated. The only reason of that can be push_back().

06:10 < lozhnikov> since it is the only line where you modify the temp vector

06:12 < jeffin143> BT , printing the memory location : says that it never changes it's memory

06:15 < lozhnikov> jeffin143: No, actually the location changed every time. (Look at the output o std::cout<<i<<" "<<&i<<std::endl;).

06:15 < lozhnikov> jeffin143: Looks like the vector reaches the maximum capacity, reallocates the array and copies the data.

06:16 < jeffin143> Omg , yes you are right, just noticed it , I was checking the last line , which I wrote

06:16 < jeffin143> Yes that's the issue

07:47 < jenkins-mlpack2> Project docker mlpack nightly build build #341: STILL UNSTABLE in 3 hr 33 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/341/

08:21 < xiaohong> Does anyone compile the mlpack with the ouput says that

08:22 < xiaohong> mlpack/build/deps/ensmallen-1.15.0/include/ensmallen_bits/spsa/spsa.hpp:120:8: warning: private field 'shuffle' is not used [-Wunused-private-field] bool shuffle;

08:26 < favre49> xiaohong: Yup, that is issue #114 on the ensmallen repo

08:32 < xiaohong> Thx, I will subscribe the issue to see the status.

12:07 < jeffin143> rcurtin : you there ?

12:10 < jeffin143> Um, regardin scaling PR , If we have to provide inversetransfrom () function to user , then we must find inverse of a matrix , and that could be risky

12:12 < jeffin143> Suppose : output = whiteningMatrix * output , now insides the inversetransfrom () , we can easily return inv(whiteningMatrix) * output , but there may be case when whiteningMatrix is singular matrix , in that case, our function wouldn't be able to do the intended job*

12:43 < zoq> xiaohong: The warning was fixed in 1.15.1.

12:49 < xiaohong> Zoq: So what I need to do is to clean builded dir, and recompile to download new version of ensmallen?

12:51 < zoq> Since this is just a warning, you could ignore the message; but yeah you can remove the build folder and the cmake file will download the latest stable version if you build mlpack again.

12:52 < xiaohong> Okay, thank you for your information.

12:53 < xiaohong> Also any information about the hoeffding_tree_test.cpp's warning?

12:54 < zoq> have to check the issue

12:55 < xiaohong> like the issue I opened in https://github.com/mlpack/mlpack/issues/1911

12:56 < zoq> right, saw the issue, not sure yet if this is a correct warning

13:58 < rcurtin> jeffin143: I have not thought it through fully (please check my math...) but the whitening transformation we are doing looks like this: output = diag(1 / (eigenvalues + epsilon)) * eigenvectors * data

13:58 < rcurtin> inverting this should be possible, but let's first consider the case where all eigenvalues > 0

14:00 < rcurtin> in this case, we can invert as: inv(eigenvectors) * diag(eigenvalues) * output

14:00 < rcurtin> and I think there is an identity that lets us simplify inv(eigenvectors) (also, the eigenvectors should be full-rank and thus invertible unless I am mistaken)

14:02 < rcurtin> I think this will work even when eigenvalues are 0, because of the '+ epsilon'

14:03 < rcurtin> now, in cases where we have negative eigenvalues, I am not sure, I think the inversion formula will still work but we may not get a unique result (I have not thought it through)

14:08 < jeffin143> rcurtin , the scaling_test file has a dataset for which eigenvectors we're [ 1,0 ; 0,1]

14:09 < jeffin143> Which can not be inverted.??

14:10 < rcurtin> that's the identity matrix, its inverse is itself

14:10 < jeffin143> I do agree but using function inv(a) throws an error

14:11 < jeffin143> Since d is 0 for this matrix

14:11 < jeffin143> Armadillo inv()

14:11 < rcurtin> I don't know what d is

14:12 < jeffin143> determinant

14:12 < rcurtin> the determinant of the identity matrix is 1

14:14 < jeffin143> Oh yeah , then I don't know , why it throwing error , just a sec

14:16 < rcurtin> you might want to look into identities you can use... remember that the eigendecomposition is A = QVQ^-1, so Q (the eigenvectors) have to be invertible

14:16 < rcurtin> I am not sure if there are any identities that apply here... maybe we have to invert Q... but it should be possible regardless

14:17 < jeffin143> Yes , eigenvectors have to be invertible , we can be sure of that

14:18 < jeffin143> I need to rework , on the code , may be it's something else that is throwing error

14:18 < jeffin143> Will get back to you in some time :)

14:18 < rcurtin> sounds good

14:19 < rcurtin> like I said I think in some cases where the given matrix has zero or negative eigenvalues anyway, we may need special handling and may not be able to recover the exact original matrix (I *think* we can recover it under some transformation but I haven't thought about it well enough), but that's okay

14:19 < rcurtin> we will still be able to match the API of the other scalers, I think, and that helps with consistency for users

15:10 < rcurtin> I'm presenting mlpack today at the DataTech conference: https://minneanalytics.org/datatech/

15:11 < rcurtin> I understand the audience will not be very technical, but we'll see how much C++ they are ready for :)

15:12 < zoq> nice speaker list

15:14 < rcurtin> unfortunately the only talks on the schedule that I find interesting are at the same time as my slot...

15:14 < rcurtin> I'll just be reusing the slides for the past 10 mlpack talks I've given... I'm surprised at the mileage I have gotten out of these

15:14 < rcurtin> especially since I originally wrote that talk in 2016

15:16 < zoq> the title is great :) I think it's the same as well

15:18 < zoq> "Leadership in Data Science Panel" sounds interesting

15:21 < rcurtin> I think I will miss that, I am still at my company's little Minneapolis office :) I was going to head to the conference at lunchtime

16:56 < jeffin143> rcurtin : sorry for the trouble , I would have certainly done some typos, that raised error, while rewriting the code , it work well enough.

16:59 < jeffin143> So I guess , we are correct :)

17:01 < jeffin143> Also good luck for your talk ;)

17:01 < jeffin143> :) *

18:17 < rcurtin> jeffin143: thanks, I'll take a look at the updated code when I have a chance :)

20:53 < sreenik> rcurtin: How was the response for your presentation?

22:32 < rcurtin> sreenik: I think it went really well, the room was not big but it was pretty full and there was a good amount of engagement and interest

22:33 < rcurtin> mostly I hear that people are scared of C++ (and rightfully so---the error messages can be insane), so getting people to write C++ instead of Python can be tricky

22:33 < rcurtin> (however, I think the Python bindings help with this for the average user. the user who is interested in speed is probably writing C++ anyway and is not scared)

22:33 < rcurtin> as always, people are excited about the Bandicoot slide, so it reminds me that I really need to be spending my free time working on that :)

22:34 < rcurtin> here are the slides I used, but they're very similar to other slides I've used in the past: http://www.ratml.org/misc/datatech_mlpack.pdf

22:34 < rcurtin> I gave out a bunch of stickers too, maybe someday they will be common on laptops :)

22:42 < rcurtin> I heard more "wow, I want to go give this a try" than I usually do, and also some "I'd like to contribute to machine learning libraries, maybe this is a good place to start", so I wonder if we will see some contributions from Minnesota soon

22:42 < rcurtin> but not Antarctica or space, which are the two places I am interested in receiving contributions from, now that we've checked off all the other continents :)