16:56 < zoq> nilay: Hello, Y are the responses (labels), the first dimension represents the number of classes and the second dimension the number of samples. So, if you test the network with a single input (3 classes) it should look like: arma::mat Y = arma::zeros<arma::mat>(3, 1); instead of arma::mat Y = arma::zeros<arma::mat>(3, 192);.

17:30 < zoq> nilay: Okay, the issue is fixed in: https://github.com/mlpack/mlpack/commit/ccc6ff0c01d670536dca496ebdf74d78497020f6

18:01 < travis-ci> mlpack/mlpack#1213 (master - ccc6ff0 : Marcus Edel): The build was broken.

18:01 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/81c14d9ec825...ccc6ff0c01d6

18:01 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/145063537

20:45 < nilay> zoq: hi, how much time (estimated) should training a net in inception_layer_test.cpp should take??

20:46 < zoq> nilay: That depends on the input size.

20:47 < nilay> input size in file is 28 x 28 x 192

20:48 < zoq> nilay: Which is one sample right?

20:49 < nilay> yes

20:51 < zoq> Since it's a single input, it should only take a couple of iterations < 20.

21:07 < nilay> ok i don't know why it just gets stuck in the forward pass only

21:08 < zoq> In the first forward pass?

21:08 < nilay> zoq: do you see any discrepancy https://github.com/nilayjain/mlpack/blob/inceptionlayer/src/mlpack/methods/ann/layer/inception_layer.hpp

21:08 < nilay> yes

21:09 < nilay> test : https://github.com/nilayjain/mlpack/blob/inceptionlayer/src/mlpack/tests/inception_layer_test.cpp

21:11 < zoq> it never returns?

21:21 < nilay> no i mean only the forward pass starts and ends, backward pass never starts

21:22 < nilay> zoq: could this be a problem with the test i wrote?

21:24 < zoq> yes nice catch, you need to put another layer in front of the inception layer, because you need the backward pass to calculate the error for the next layer, but there isn't a next layer.

21:24 < nilay> i did put a convLayer

21:25 < nilay> https://github.com/nilayjain/mlpack/blob/inceptionlayer/src/mlpack/tests/inception_layer_test.cpp

21:25 < zoq> yeah after the InceptionLayer right?

21:26 < nilay> yes as you can see in the test

21:26 < zoq> yes, so I was talking about putting another laer before the InceptionLayer (in)

21:26 < zoq> *layer

21:27 < nilay> oh ok let me do that then

21:28 < zoq> It could be a conv layer with a 1x1 kernel or a linear layer.

21:30 < zoq> A conv layer would be a good idea.

21:31 < zoq> or wait, BaseLayer<IdentityFunction, arma::cube, arma::cube>, should also work

21:40 < nilay> for base_layer Datatype& error would always be a matrix?

21:42 < zoq> This is my test: https://gist.github.com/zoq/cfbd06a91d722dab33ae058f271b5034.

21:42 < zoq> It calls the backward function, since there is some error in the backward pass, I removed everything in the backward pass and just added a print statement.

21:43 < nilay> yes, don't know why that error wasn't coming before?

21:43 < zoq> because the backward function of the inception layer was never used

21:45 < nilay> ok