Crashing on training with binarized values

When training with  MNIST dataset binarized to values 0 and 1 (typed Int) it crashes. 
Works if it is real valued to 0.0 and 1.0.

The issue occurs for training generally with categorical data.

The error is in forward pass function, due to inexact conversion exception thrown, when converting real value into an int.
ex. Int(1.23)
