Closed pencilzhang closed 3 years ago
@pencilzhang Hi, thank you for appreciating our work. We follow the practice of most BIQA papers to split the training/testing test by an 80%/20% ratio. So, there is no data reserved for validation. For a more strict evaluation without referring to the results on test sets, we can always use the latest checkpoints for evaluation, although we see no big difference between both strategies.
@zwx8981 Thanks for your response!
Thanks for your great job! Did you use validation data to obtain the best checkpoint (model)?