Training with Mixed-Precision Floating-Point Assignments