I have a question, I found that ISIC2018 dataset uses all the Train and Val datasets on the official website for training, and uses Test for verification and testing, which makes the value of the experimental results appear higher. However, if you only use Train for training and Val for verification, you will get a poor result if you only use Test for verification once after the training is completed. However, the former training method seems unreasonable, because the test should be invisible to the model during training, and this way is to tune parameters on the test.（translated by google）
我有一个问题，我发现ISIC2018数据集使用了所有官网中的Train和Val数据集做训练，使用Test作验证和测试，这种训练方式会使实验结果的数值看上去更高。但是如果仅使用Train做训练，Val进行验证，训练结束之后仅使用Test进行一次验证就会得到一个较差的结果。而前者这种训练方式似乎不太合理，因为Test应当在训练中对模型是不可见的，这种方式是在Test上进行调参，无法完全体现出模型的泛化性能。
I have a question, I found that ISIC2018 dataset uses all the Train and Val datasets on the official website for training, and uses Test for verification and testing, which makes the value of the experimental results appear higher. However, if you only use Train for training and Val for verification, you will get a poor result if you only use Test for verification once after the training is completed. However, the former training method seems unreasonable, because the test should be invisible to the model during training, and this way is to tune parameters on the test.（translated by google） 我有一个问题，我发现ISIC2018数据集使用了所有官网中的Train和Val数据集做训练，使用Test作验证和测试，这种训练方式会使实验结果的数值看上去更高。但是如果仅使用Train做训练，Val进行验证，训练结束之后仅使用Test进行一次验证就会得到一个较差的结果。而前者这种训练方式似乎不太合理，因为Test应当在训练中对模型是不可见的，这种方式是在Test上进行调参，无法完全体现出模型的泛化性能。