cifar10中的pytorch随机标签实验有什么用？

Question

我找到了以下代码：

DELETE FROM invitations AS inv, users AS us
WHERE inv.user_id = us.user_id
AND us.role = 0 OR 2 OR 3
AND inv.invited_at < NOW() - INTERVAL 3 DAY;

对数据集进行混音的确切程度如何？是在每次采样数据集时保存它还是翻转标签？还是先保存所有随机标签？如何运作？