Question about the naive baselines about lae

[gqk]

Hi klauscc,

I'm very sorry for the delayed response as I've been quite busy lately.

Thank you for your interest in our work.

Regarding Seq-FT, we use the same training and inference method as DualPrompt, except for the use of SGD optimizer, learning rate of 0.01 (scale based on batch size 128), and weight decay of 5e-4, since we fine-tune all parameters of the pre-trained model. The difference in performance could be attributed to the local classification loss.

Regarding the naive baseline, your understanding is basically correct. We did insert the prefixes with length 20 into each of the shallowest 5 transformer blocks and used the same training and inference method as DualPrompt. You can verify this using the code released by Jaeho Lee or the DualPrompt authors.

If you have any further questions, please don't hesitate to reach out to me. And if I don't reply promptly, please feel free to send me an email (gqk at stu.pku.edu.cn) as a reminder.

Thank you again for your interest.

Best regards,

Qiankun Gao

[gqk]

Hello @klauscc,

I noticed that there hasn't been any progress on the issue for quite a long time, so I will go ahead and close this issue. If you have any other questions or concerns, please don't hesitate to reach out to me again.

Best regards,

Qiankun Gao