00:58 < tham> nilay : Hi, IMHO, you can ignore all of the optimization issue right now

00:59 < tham> Make your codes robust first before you think about optimization

01:00 < tham> Worry about optimization before the algorithm complete is too early

01:01 < tham> I remember Herb Sutter talk about what are the difference between immature optimization vs pessimistic

01:04 < tham> In short(for what I understand), immature optimization means you try to make the codes become more complicated without measuring this

01:05 < tham> is worth or not(can boost up the performance as you expected or not)

01:07 < tham> Another case would be consider as immature optimization is you try to optimized the codes before it is stable

01:07 < tham> What is pessimistic?

01:08 < tham> pessimistic means you prefer more expensive solution to solve the problem when there are cheaper(faster, smaller memeory etc) and

01:09 < tham> simple(would not increase complixity of codes, like swap vs copy, prefix vs postfix) solution exist

01:14 < tham> If you already finish the algorithm, unit test return perfect results

01:14 < tham> Then I think I am worry too much

01:15 < tham> Whatever, I think you can postpone the optimization task until all of the algorithms are done(include googleNet)

01:16 < tham> Roman was not build in one day

02:31 < tham> Forgot to mentioned, it is a good spirit to consider optimization for the algorithm

03:17 < tham> but we could wait until the codes are stable, tested

04:42 < nilay> tham: ok, i get it.

07:49 < nilay> zoq: Hello, can you tell me what does the pos_loc, neg_loc block does, in prepare_data() function. And if there is an easier way to do it..

09:54 < zoq> nilay: pos_loc are the locations (x, y) of edges and neg_loc are the locations of nonedges. Take a look at the boundary parameter, it's a binary matrix where 1 defines an edges and 0 respectively a non-edge.

09:55 < zoq> nilay: So, if you train a tree you like to get a normal distribution over non-edges and edges. Since the python code uses a parameter to define the split between pos_loc and neg_loc it's kind of complicated (e.g. 80% pos and 20% neg). The default value is 50/50, so I think, it's valid to say we implement the 50/50 split only.

09:55 < zoq> nilay: So what you could to is to iterate through the boundary and use ind2sub as long as you got x pos samples and x negative samples.

10:28 < nilay> i cannot compile with ind2sub do i need a new armadillo version?

10:38 < zoq> nilay: ah, sorry I thought it would be available in 4.1 but it was introduced in 6.4.

10:47 < zoq> nilay: https://gist.github.com/zoq/89c694907c2b4c98d00a6aac87911d4f

11:21 < nilay> zoq: is the function you wrote correct? you're not using rows anywhere. and from docs i don't get an idea what does ind2sub do.

11:37 < zoq> nilay: Yes, I'm sure here is the armadillo function: https://gist.github.com/zoq/b668fd75af14cd710ba2085703710dc2, you should use row instead of col, I get that wrong. Since its continues, you only need the number of columns or rows.

11:37 < zoq> nilay: So what this function does is to map from an linear index to subscript: http://de.mathworks.com/help/matlab/ref/ind2sub.html

11:39 < zoq> nilay: So you have this boundary matrix, you search for a 0 or 1 and would like to get the subscript (x, y) to generate the location. I think you could also use two for loops for that.

11:43 < nilay> zoq: in armadillo we will count linear indices in column major order?

11:43 < zoq> nilay: yeah, right

12:01 < nilay> how do we find labels using this loop: http://pastebin.com/PnKa3UkC

12:15 < zoq> nilay: So what this loop does, is to iterate through all locations and extract the label from the segmentation matrix by using the location as the center and the patch size.

12:15 < zoq> nilay: The segmentation is given, (small_segmentation_1.csv), take a look at the example segmentation image I created. It's basically a matrix with the size of the input image in which each pixel is assigned with a segmentation label.

12:15 < zoq> nilay: So what, you should do is to load the segmentation and extract the label using the location and the defined patch size.

12:16 < zoq> nilay: It's 101085_segment.png

12:17 < zoq> nilay: The small_segmentation_1.csv file contains two segmentations for two images

12:20 < nilay> zoq: thanks, i understand now.

12:23 < zoq> nilay: okay, great