Hi, I was working on developing models and whenever I exceeded a certain limit of output maps in convolution layer, and then tried to train, I got the error : Segmentation fault (core dumped)
basically:

model.Add<Convolution<>>(3, 16, 7, 7, 2, 2, 3, 3, 28,28);
model.Add<Convolution<>>(16, 192, 3, 3, 1, 1, 1, 1, 10,10);
...
model.Train(Xs,Ys,optimizer);

works, but

model.Add<Convolution<>>(3, 64, 7, 7, 2, 2, 3, 3, 28,28);
model.Add<Convolution<>>(64 192, 3, 3, 1, 1, 1, 1, 10,10);
...
model.Train(Xs,Ys,optimizer);

terminates with error:

Segmentation fault (core dumped)

As suggested by @zoq, I used gdb to find what exactly is causing error and I found that

arma::Mat<double>::operator=(arma::Mat<double> const&) ()

inside Gradient() function of Convolution<> class is causing this error when called during Training process, which probably means when a matrix is being filled, but where exactly I have no idea.

Full backtrace: https://pastebin.com/bvVzJJjs

I'm not able to figure out further what to try.
Any help/suggestion are welcome.