Cross entropy loss is divided by the number of classes and not the batch size

[cDenius]

In the current cross entropy implementation, the loss value is divided by N. In principle, this should be the batch size. But since we are working with a batch size of one and thus call the function with a one-dimensional array, we are currently dividing by the number of classes which makes no sense at all.

Deleting the division would solve the problem for batch size one (= dividing by one). But to support different batch sizes in the future, a more sophisticated solution would be nice.

[eileen-kuehn]

@all-contributors please add @cDenius for bug

[allcontributors]

@eileen-kuehn

I've put up a pull request to add @cDenius! 🎉