12:10 < alexsc> rcurtin: went for the argmax, 4 lines in armadillo. let me know if I shall integrate it somewhere!

12:11 < alexsc> + my implementation of KNN gives the choice of weighting window among the K neighbors.. rectangular (same weight to all K), triangular, and so on

12:11 < alexsc> is there any way to achieve this currently in mlpack?

12:12 < alexsc> I can answer myself.. nope :). I’ll code it

13:52 < rcurtin> alexsc: yeah, unfortunately, there is not an existing way to do that

13:52 < rcurtin> if you implement it, though, it would be great to include the functionality, I would be happy to work with you on a PR to get it merged in

13:52 < alexsc> would you see it as a template arg?

13:53 < rcurtin> since we have so many techniques that are related to nearest neighbor search, I'd personally think it should be a standalone function that you put your neighbors/distances results into

13:53 < alexsc> enum KNNWeightingKernelType {

13:53 < alexsc> kRectangular = 0, // 0.5 · I(|d| ≤ 1)

13:53 < alexsc> kTriangular, // (1 − |d|) · I(|d| ≤ 1)

13:53 < alexsc> kEpanechnikov, // 0.75 * (1 − d^2 ) · I(|d| ≤ 1)

13:53 < alexsc> kQuartic, // 15/16 * (1 − d^2)^2 · I(|d| ≤ 1)

13:53 < alexsc> kInversion // 1 / |d|

13:53 < alexsc> };

13:53 < rcurtin> alternately, if you want to do something more complex, I think you could make a class like

13:53 < alexsc> would be worth it, normally accuracies are so much better with Triangular windows

13:53 < rcurtin> class KNNClassifier<template KNNSearchType>

13:54 < rcurtin> and then inside of that KNNClassifier class, you would just call KNNSearchType(k, neighbors, distances) and then use KNNWeightingKernelType to figure out the labels

13:54 < rcurtin> this would be a really nice contribution if you were willing to take the time :) (and one that probably should have been done years ago... I'm sorry the support you need isn't already there)

13:56 < alexsc> cool, will do it beginning of the next week… I have a deadline on friday

13:56 < rcurtin> sounds great, and let me know how I can help out if you need. good luck with the deadline :)

13:59 < alexsc> sure, thanks for now!

14:00 < alexsc> that NCA is making wonders… +30% accuracy over untreated KNN

14:18 < alexsc> question: sorry for bothering today

14:19 < alexsc> I compiling my KNN on mac is ok. now on Ubuntu, installed with apt-get, I am getting index_max not a member of arma in RcppArmadillo

14:19 < alexsc> sorry, bad paste from stackoverflow: I get arma::uvec {aka class arma::Col<long long unsigned int>}’ has no member named ‘index_max’

14:19 < alexsc> do you know how to update armadillo to a more current version?

14:47 < alexsc> k. built from source :)

14:50 < rcurtin> oops, sorry, stepped out for a few minutes

14:51 < rcurtin> index_max() is available from 7.300 onwards, so in ubuntu 14.04 or 16.04 the version is too old

14:51 < rcurtin> so yeah, compiling from source is probably the best option

21:58 < zoq> rcurtin: Having fun with #1125 and travis.

22:10 < rcurtin> I wouldn't call it fun... :)

22:10 < rcurtin> I'm flabbergasted first that ubuntu 16.04 has a background auto-update process, and second that the travis images don't automatically disable it and prevent it from running

22:11 < rcurtin> I think that even though I run 'systemd stop apt-daily.service apt-daily.timer', the process has already started by then and continues, so I may have to find some other way to kill it or prevent it from starting...

22:14 < zoq> hmm, I am confident that you solve the issue

22:16 < rcurtin> yeah... I don't know how long it will take though...

22:16 < rcurtin> I think I will have to find the travis-ci docker images and run them directly, but I haven't taken that step yet

22:16 < rcurtin> really I would rather be writing code instead of debugging CI tools :)

22:17 < rcurtin> but such is the nature of CI :)

22:21 < zoq> We don't use the travis docker container, we use the VM on GCE, the last time I checked running on GCE was way faster; besides GCE provides 7.5GB of RAM instead of 4GB.

22:23 < rcurtin> ah, ok; I hadn't looked into it much

22:23 < rcurtin> I do suspect there may exist a docker container with the exact same configuration as GCE, but let's see what this round of debugging brings back before I look into that...

22:24 < zoq> right, maybe there is an docker image