14:07 < nilay> rcurtin zoq: in the conv2 function of armadillo, are the boundaries padded with 0, or are the boundaries reflective?

14:09 < nilay> can i use this function to do 2D convolution, regardless of how boundaries are considered (does it matter much whether we pad 0 boundary or reflective boundary)?

14:14 < nilay> also we will have to backport this function if we want to use it, so is it worth it or i just use the Naive convolution class implemented in mlpack

14:17 < zoq> nilay: It's zero padded, also the padding shouldn't really matter.

14:17 < zoq> You are right, better you use one of the Convolution rules from the ann modules here.

14:18 < nilay> zoq: we need output "same"

14:18 < nilay> not full or valid mode

14:20 < nilay> i could not find the constructor for "same"

14:20 < nilay> in arma they have full and same, so i can use same

14:21 < nilay> or i can leave it as it is

14:21 < nilay> i have used 2 loops and submat which seems decent enough?

14:22 < zoq> You could just use full and use submat to get the same size as the input data, that's what arma does.

14:22 < zoq> Do get the gradient?

14:22 < zoq> *To

14:24 < nilay> to perform triangle convolution, i made a SepFilter2D function and you suggested to use Naive Convolution

14:25 < zoq> let me take a look

14:26 < nilay> so i used 2 loops and submat. (but i also did a copymakeborder to make borders) and naive convolution uses 4 loops and pointers, it seems. i am not sure

14:31 < zoq> hm, I'm not sure, which one is faster, do you think it's reasonable to test it with some inputs? I wrote the comment because the convolution method is well tested and I'm working on a Winograd based convolution method that should be faster for small kernels.

14:39 < nilay> yeah sure. so i'll give an input to both and use time command, will it do the job?

14:40 < zoq> yeah, that should be good enough.

14:40 < nilay> ok

15:13 < zoq> nilay: Can you open a PR or update the existing, once you finished the mapping function?

15:14 < nilay> zoq: yeah sure

15:15 < nilay> i am in the process of doing it. it is my first priority. sorry for the delay.

15:15 < nilay> also, i am trying to install arma 7.2

15:15 < nilay> i already had arma 7.something

15:16 < zoq> nilay: okay, great, no need to do it today or tomorrow

15:16 < nilay> do i first have to remove it and then install 7.2

15:16 < nilay> i think we can make use of index_min and index_max introduced in 7.2 . I will backport them for mlpack

15:17 < zoq> you can just build against the specific version with: ARMADILLO_INCLUDE_DIR=....

15:18 < zoq> and ARMADILLO_LIBRARY=...

15:22 < nilay> ok yeah. i remember those now. But still if i don't want to give these link commands and only use the recent version, can i do something about it?

15:23 < zoq> remove the old version and install the version you like to use to build mlpack

15:28 < nilay> ok thanks, and one more thing, can i download the irc logs on my computer and grep it if i want. It is sometimes hard to remember the date on which i was told something.

15:29 < nilay> i mean i can do this automatically writing some python code, i was asking if i have an easier way

15:30 < zoq> maybe ryan, can provide the full log file

15:30 < nilay> that would be great

17:11 < rcurtin> nilay: hang on, let me get the raw log

17:12 < nilay> ok

17:15 < rcurtin> nilay: http://ratml.org/misc/mlpack.log

17:15 < rcurtin> it is kind of long, but hopefully it is helpful

17:15 < rcurtin> maybe I should add an option to the IRC webpage to download the raw log, but I don't have time to do that today

17:19 < nilay> rcurtin: thanks, this helps a lot :)

19:09 < TD> I think the main error that I am running into is setting up the project in VS2015. I tried to build mlpack_cf and I received LNK1104 cannot open file 'kkernel32.lib'

19:10 < TD> Is there any documentation about how set up a solution / project for VS2015? Properties etc?