Closed Forrest-110 closed 8 months ago
Hey @Forrest-110, apologies for the very late reply I must have missed this. If you are still working on this, do you notice a similar trend with your downstream performance when doing inference? Typically you would only want to use soft rank/sort while training and the gradients are needed. Once trained, you can just use torch.argsort
as you do above. I would recommend tuning the regularization using a cross validation set.
这是来自QQ邮箱的自动回复邮件。 您好,邮件已收到,但我无法亲自回复您的邮件。不久即速回。
这是来自QQ邮箱的自动回复邮件。 您好,邮件已收到,但我无法亲自回复您的邮件。不久即速回。
Dear authors, you have done a great job! But I met some difficulties when dealing with large size of input (1k-5k).
I have a list of numbers ranging in [0,1], and the list size is about 1k to 5k. And I'm trying to find a suitable regularization strength to get accurate ranking results.
My experiment codes are listed below
The things is:
That's exactly the issue. If I have to find a regularization strength for every input size, it could be quite impracticable. Could you help me?