Adaptive Gradient Methods with Dynamic Bound of Learning Rate