Not at all! Every change in hyperparameters, every run of training on our data where we use our validation set to verify results, makes it more plausible that we will overfit to our validation set! In this case, the overfitting would come from picking an architecture, picking training params, that perform well on our validation set but that do not generalize to other unseen data.

To be rigorous about this, we should set aside a third bit of data, called the test set. This is a part of our dataset that is neither used during the training nor for calculating our metrics. We do not look at this until the whole project is finished