Towards best practice in explaining neural network decisions with LRP