Fast generalization error bound of deep learning without scale invariance of activation functions