Calculating accuracy by not using test data
Why you are not using test data instead of validation data to calculate an accuracy? If you are not using test data, why did you divide it at the beginning?
I believe it is because in this stage we are still trying to find the best hyperparameters. Therefore, we can still change the architecture, etc... The test dataset is the final evaluation, when everything is already set and we want to find the final model performance. After evaluating with the test data we "cannot" change anything else.