Closed Theheavens closed 3 years ago
In Sections 7.3 and 7.4, the performance used in ranking analysis are all validation performance. First, the validation performance in the final epoch meanss it will run out of all epochs, which is final epoch, am I right?
Yes, that is how the experiment was done in our "Design Space for GNN" paper. The main reason is that we wanted to inspect the effect of training epochs as well. If early stopping is use, the training epoch parameter does not make a difference.
Second, I am wondering why we don't use the early-stop and use the test performance mentioned below, which is the best validation epoch test performance.
In practice, we should focus on the epoch where the validation performance is the best. This is conveniently provided in GraphGym in "test_best.csv".
I see that 7.2 Experimental Setup in paper
In Sections 7.3 and 7.4, the performance used in ranking analysis are all validation performance. First, the validation performance in the final epoch meanss it will run out of all epochs, which is final epoch, am I right? Second, I am wondering why we don't use the early-stop and use the test performance mentioned below, which is the best validation epoch test performance.