14:20 < nilay> zoq: i am having problem in understanding numpy.dot

14:20 < nilay> this is because numpy and arma seem to store arrays as different

14:21 < nilay> a (4,4,3) array in numpy is four, (4 * 3) matrices

14:21 < nilay> and in arma it is three, (4 * 4) matrices

14:22 < nilay> so the code i have written for this line: xyz = N.dot(src, rgb2xyz_mat) , in the RGB2LUV function is being incorrect

14:23 < nilay> and i am unable to fix it, having spent quite a lot of time on this.

14:38 < zoq> numpy is row major and armadillo is column major, so everytime numpy iterates over the rows you iterate over the columns

14:38 < zoq> another solution is to transpose the matrix, so you could do src.t() or arma::trans(src).

14:39 < zoq> Does, this solve the problem?

14:41 < nilay> src is a 3d matrix

14:41 < nilay> a cube

14:42 < zoq> and I guess rgb2xyz_mat is a 2d martix?

14:42 < nilay> yes

14:43 < zoq> and the output of xyz is a 2d matrix?

14:44 < nilay> 3d matrix

14:46 < zoq> okay, so what you do is to iterate over the slices of src: xyz.slice(i) = src.slice(i) * rgb2xyz_mat?

14:46 < nilay> no

14:46 < nilay> src.slice(i) is of dimension (img.rows, img.cols)

14:46 < nilay> rgb2xyz_mat is of dimension (3, 3)

14:47 < zoq> and img.rows is the actual image size?

14:47 < nilay> yes

14:48 < zoq> hm, sounds weird, I need to take a look

14:48 < nilay> documentation of numpy.dot says: For N dimensions it is a sum product over the last axis of a and the second-to-last of b. so i tried writing this using loops, but then also i can't get correct answer

14:49 < nilay> if we are doing numpy.dot(a, b)

15:12 < zoq> okay, so what this does is to transform the rgb values into the CIE 1931 XYZ values.

15:13 < zoq> So, you multiply the rgb values with the transformation matrix. Right now I can only think of a naive implementation.

15:18 < zoq> https://gist.github.com/zoq/13eb9ea35b71419c38017066071deaf2

15:18 < zoq> should be double b = src.slice(2)(i);

15:47 < nilay> is this ( double b = src.slice(2)(i) ) syntax correct?

15:48 < nilay> ok it should be vector

15:49 < zoq> hm, not sure, I thought it would work? It should return a double

15:50 < nilay> src.slice() is a matrix

15:50 < nilay> since src is a cube

15:50 < nilay> so matrix(i) should be vec?

15:50 < zoq> yes, but we like to get the element of that matrix at position i

15:51 < nilay> ok let me try.

15:53 < nilay> zoq: syntax is correct

15:54 < nilay> but the test is failing

15:54 < zoq> can you post the outut and the reference output?

15:55 < nilay> seems like the problem is numpy and arma array storage. i'll post it in a minute

16:00 < nilay> http://pastebin.com/9E0CGJh3

16:02 < nilay> this is our rgb2luv function: http://pastebin.com/s0yxyq92

16:09 < zoq> how does the test input look like?

16:09 < zoq> I mean the mlpack test input

16:11 < nilay> zoq: https://github.com/nilayjain/mlpack/blob/feature-extraction/src/mlpack/tests/edge_boxes_test.cpp

16:12 < nilay> you can run the test it'll print the input given and desired output and that test fails

16:14 < zoq> okay, I'll take a look at the code, it might take some time, I need to finish some other things first

16:14 < nilay> ok

17:20 < zoq> src = N.array([[[0.2, 0.1, 0.2], [0.1, 0.0, 0.1], [0.0, 0.1, 0.2]], [[0.2, 0.1, 0.2], [0.1, 0.0, 0.1], [0.0, 0.1, 0.2]],[[0.2, 0.1, 0.2], [0.1, 0.0, 0.1], [0.0, 0.1, 0.2]]], dtype = N.float64)

17:20 < zoq> So it turns out the function interprets the input as RGB, RGB, RGB ...

17:20 < zoq> I thought the RGB value at (0, 0) would be (0.2, 0.2, 0.2) and not (0.2, 0.1, 0.2)

17:21 < zoq> so, I guess, what you could do here is to adapt the input accordingly

17:24 < zoq> also, I think I messed up the xyz matrix, it should be:

17:24 < zoq> xyz.slice(0)(i) = 0.430574 * r + 0.341550 * g + 0.178325 * b;

17:24 < zoq> xyz.slice(1)(i) = 0.222015 * r + 0.706655 * g + 0.071330 * b;

17:24 < zoq> xyz.slice(2)(i) = 0.020183 * r + 0.129553 * g + 0.939180 * b;