10:41 < ironstark> I feel that the idea is good. I'll implement it and open up a PR

10:42 < rcurtin> in case anyone is interested in what I am up to and why I am so quiet on the mlpack Github, I have posted some pictures at http://www.ratml.org/misc/australia_pics.html :)

11:47 < zoq> ironstark: I think we should discuss first which architectures we like to compare. We could start with a simple two layer multilayer perceptron with e.g. 512 nodes in the first layer and 256 nodes in the second, and use sigmoid as activation function. For training we use SGD with momentum, it looks like that's all you can do right now with dlib.

11:53 < ironstark> zoq: okay I will use this architecture.

11:53 < ironstark> rcurtin: Looks like a really fun trip :)

22:21 < ironstark> zoq: rcurtin; This is my dlibml SVM script

22:21 < ironstark> https://paste.ubuntu.com/25242977/

22:21 < ironstark> I am getting all the assignments as 0 here

22:23 < ironstark> never mind, got the mistake. Code works now

23:00 < ironstark> actually its breaking again. Same error. All the predictions are coming out to be 0

23:00 < ironstark> https://paste.ubuntu.com/25243160/

23:00 < ironstark> Can you please look into this

23:10 < zoq> ironstark: What dataset did you use for the test?

23:10 < ironstark> i used the iris dataset. I have tried to implement this as a multiclass classification svm

23:11 < ironstark> Have i made a mistake in doing that?

23:20 < zoq> sample_type m2;

23:20 < zoq> m2.set_size(trainData.n_cols);

23:20 < zoq> for (size_t j = 0; j< trainData.n_cols; ++j)

23:20 < zoq> {

23:20 < zoq> m2(j) = trainData(trainData.n_rows - 1, j);

23:20 < zoq> labels.push_back(m2);

23:20 < zoq> }

23:20 < zoq> Is that the part where you put the labels together (the last row holds the label)? It looks like m2.set_size(trainData.n_cols); isn't correct since for me the label is a single element so I would expect m2.set_size(1); maybe I missed something?

23:23 < zoq> sumedhghaisas_: Any updates regarding the gradient check? Still going through the code ...

23:32 < kris1_> lozhnikov: On 5000 images with training set of size 4000 and test size 1000 images the accuracy is around 30% for poolSize = 4 lr rate 0.01(time 20 min) poolSize = 3 lr rate 0.3 accuarcy is 90% 21 min

23:33 < kris1_> Have a look i have updated the PR.

23:33 < kris1_> I think will investigte why there is so much of diffrence with changing the parameters so litlle

23:36 < kris1_> zoq: Were you able to look at the GAN. Were you able to find any obvious errors !!

23:42 < zoq> kris1_: Glanced over the PR, but couldn't see anything obvious, trying to do an in depth look over the weekend.