Training, validationa and test sets

A test set is meant to be only used once to evaluate your model. Making any changes to the model based on performance results on the test set means the test set is now contaminated, i.e. when used in the learning process in any manner it is not a test set now. [Paraphrased from here]