hello,I really reproduced the acc in your paper,but I have some questions.
the early stop you set so right,when I set it bigger,the best model in val will be better,but worse in test ,so I think,you may select the model directly by the test dataset,it may be more reasonable first find the best in val and then test,not test then set the early stop to make the training process right stop in the epoch which test best in the test.
maybe can I think this way:For the publish dataset,test is already lose its oringin function when test is public,so making it fair comparison is just ok
hello,I really reproduced the acc in your paper,but I have some questions.
the early stop you set so right,when I set it bigger,the best model in val will be better,but worse in test ,so I think,you may select the model directly by the test dataset,it may be more reasonable first find the best in val and then test,not test then set the early stop to make the training process right stop in the epoch which test best in the test.
so can you explain it,thanks