Details for reproducing

[KSXGroup]

Hi, I am trying to reproduce your result on eth dataset but it is far away from the reported result (just better than linear).

I preprocessed from raw data using your data converter and pretrained for 700 epoch (lr=3e-6) + 300 epoch (lr=3e-7) with lambda value you mentioned in paper. And then finetuned on eth. But I could not got the reported value. Maybe I missed some steps. So could you provide detail for reproducing or trained model checkpoint ?

Thanks!

[Sigta678]

Hi @KSXGroup,

Thanks for your question.
The missing steps might be the learning rate decay on the fine-tuning process, or you can try to drop some data and leave just 1% or 10% for fine-tuning. We discussed this part in our experiment "amount of data for fine-tuning". Also, you can check our Figure 2 in supplementary materials, this explains the phenomenon you met on eth.

Hope this helps!