Layer-wise Adaptive Gradient Norm Penalizing Method for Efficient and Accurate Deep Learning