05:12 < lozhnikov> kris1: which test do you mean? Could you upload that? I tried to compile this test (https://gist.github.com/kris-singh/9420d3e9a0afe12099826dd98222dce7) but the build failed. So, it seems we use different versions. Could you upload the latest code?

06:48 < lozhnikov> kris1: It seems the shape of the input is incorrect in the second convolution layer of the generator

06:48 < lozhnikov> try to use the following:

06:48 < lozhnikov> generator.Add<Convolution<>>(gInputDim / 2, gInputDim / 4, 3, 3, 2, 2, 1, 1, 28, 28);

06:48 < lozhnikov> i.e. replace 56 by 28

13:20 < kris1> Hi lozhnikov i just came home…..I have pushed the changes…. have look…… Here is the evaluation file that was giving me undefined behaviour at times.

13:20 < kris1> chttps://gist.github.com/kris-singh/9420d3e9a0afe12099826dd98222dce7

13:22 < lozhnikov> kris1: I have already found the issue

13:22 < lozhnikov> It seems the shape of the input is incorrect in the second convolution layer of the generator

13:22 < lozhnikov> try to use the following:

13:22 < lozhnikov> generator.Add<Convolution<>>(gInputDim / 2, gInputDim / 4, 3, 3, 2, 2, 1, 1, 28, 28);

13:22 < lozhnikov> i.e. replace 56 by 28

18:38 < kartik_> hi <zoq> why does that happen that.. when the vanilla network has labels named as 0 1 2 for 3 output layer nodes

18:38 < kartik_> matrix build error occurs

18:38 < kartik_> but for 1 2 3 as nomenclature for output nodes

18:38 < kartik_> it runs fine

18:40 < kartik_> <zoq> iris dataset is working fine and also converging very fast

18:40 < kartik_> after i change the labels to the above names

19:10 < zoq> kartik_: hm, not sure right now

19:11 < zoq> kartik_: build error or assertion?

19:12 < kartik_> for the iris dataset

19:12 < kartik_> all i do is 4 input 4 hidden and 3 output

19:12 < kartik_> then on the training labels marked by 0/1/2

19:12 < zoq> What is the actual error you see?

19:12 < kartik_> i just add one to everything 1/2/3

19:13 < kartik_> and now the network works fine

19:13 < kartik_> error: Mat::operator(): index out of bounds terminate called after throwing an instance of 'std::logic_error' what(): Mat::operator(): index out of bounds Aborted (core dumped)

19:14 < zoq> kartik_: Can you use MeanSquaredError instead of NegativeLogLikelihood and test again?

19:14 < kartik_> <zoq> but its working fine and the run time has got significantly low

19:15 < zoq> kartik_: Sounds good, we should still see if we can figure out what the actual problem is.

19:20 < kartik_> error: subtraction: incompatible matrix dimensions: 3x1 and 1x1 terminate called after throwing an instance of 'std::logic_error' what(): subtraction: incompatible matrix dimensions: 3x1 and 1x1 Aborted (core dumped)

19:20 < kartik_> this is the new error i get

19:21 < kartik_> <zoq> here is the code https://ghostbin.com/paste/ueesx

19:24 < zoq> I see, the problem here is that if you use the MeanSquaredError layer the label size has to match with the output layer size, in yur case the label size is 1 and the network outout size is 3

19:24 < kartik_> with mean square error .. even uncommenting line 25 and 29 gives error which works fine in negativelikelihood

19:28 < zoq> kartik_: You need to preprocess the labels first something like: https://gist.github.com/zoq/0f1651780edeec49391b1bfb1ae46532

19:30 < kartik_> https://ghostbin.com/paste/jg3h5

19:30 < kartik_> <zoq> this works fine for me.

19:30 < kartik_> should i use the MSE layer

19:30 < kartik_> or this will do for the test case

19:32 < zoq> hm, the + 1 is strange, I'll see if I can figure out what happend here, I guess if the MSE layer works the problem is the NLL layer.

19:36 < kartik_> ohkay.. thanks :)