12:04 < jenkins-mlpack2> Project docker mlpack nightly build build #152: STILL UNSTABLE in 6 hr 50 min: http://ci.mlpack.org/job/docker%20mlpack%20nightly%20build/152/

12:16 < mlPackApprentice> Hello. My team and I have made steady progress on building an RNN that predicts stockprices. There's just 1 thing that's bothering me. The RNN<>.Train(predictors, responses) function takes an arma::cube of predictors and an arma::cube of responses. The predictors is our cube of training data of hourly stockprices, divided in sequences of 24 (so 24 slices).

12:20 < mlPackApprentice> For the responses cube, I understand it to work as follows: the rows should be what classes you want predicted (in our case 5 values so 5 rows). Then the columns should be equal to the predictors cube, as you want a prediction for each column (or data sample). Then for the slices is my question: If you only make your "responses" cube to have 1 slices, does that mean it will predict only the next 1 hour based on the 24 previo

12:22 < mlPackApprentice> In other words, if my "Predictors" cube is 12x100x24, does that mean that if I make my "Responses" cube 5x100x1, I will get 100 x 1 predictions of all 5 classes for each of the samples in the corresponding slice?

12:24 < mlPackApprentice> A second question: I noticed that if I initialize my "Responses" cube with "zeros" (so a matrix with only 0's everywhere), the Train() function will give me an IndexOutOfBoundsException. But if I initialize the "Responses" cube with "randu" (so values between 0 and 1), it works. So I am wondering: If the "Responses" cube is going to be filled with predictions from the RNN anyway, why should the fill values matter? They're go

12:26 < mlPackApprentice> Thank you in advance for taking the time to read.