Optimizing Explanations by Network Canonization and Hyperparameter Search