Optimize constraints #54
Description
Currently, the keras and torch constraints are expensive due to the back and forth of the weights between the native and R representations. In fact, keras shows a warning saying that the callback takes too much time compared with the rest. We should see if there's a better way to implement this without bringing the matrices to R.