Concise Multi-head Attention Models