CLabels

// summary of API ideas

//option1:(tags) no explicit instantiation of sub-class, runtime error when labels are used
// note the subclasses (CBinaryLabels, ... ) do exist still (implementing e.g. "get_binary_vector",
// or "get_confidence_matrix", they are just never instantiated by the user

labels = CLabels()
labels.set_data(file, vector, or matrix) // tags
{
 how to do this?
}

// user calls algorithm
machine.set_labels(labels) // tags, no runtime error
predicted = machine.apply()
{
  // algorithm knows which representation it wants, so can now "cast"
  // preferably, this vector would be template free
  SGVector label_vector = get("labels").as_binary().get_values() // potential runtime error
  
  // how is as_binary implemented?
}

//user extracts
SGVector predicted.as_binary().get_values() // works, returns vector {-1,1}
SGVector predicted.as_binary().get_scores() // works, returns real vector
SGMatrix predicted.as_multiclass().get_values() // works, vector {0,1}
SGMatrix predicted.as_multiclass().get_scores() // works, real matrix (with only two rows)
predicted.as_regression().get_values() // runtime error, no point to convert discrete to real

It looks like we need to define what type apply() and as_binary() should return. If we want both as_binary().get_values() and as_binary().get_scores(), in addition to label values we need to have prediction scores in there too.