Improving Equivariant Model Training via Constraint Relaxation