Shiyu Xuan comments

Results 14 comments of


                                            Shiyu Xuan

Some questions about the appearance space and the structure space

> Hi @BossBobxuan , @Zonsor > Sorry for the late response. Yes. It is possible. > > However, we did not do it. The main reason is that the structure...

Question about inference

Thanks for your kindly replying!

训练过程中otb的精度

谢谢你了

训练过程中otb的精度

我自己写了一个训练代码，但是结果一直比较诡异，现在问题解决了，谢谢你了，我感觉用youtube bb训练在OTB能提升3-4个点

Question about the training strategy in the stardard FSL setting

I have the same question with you!

Is the AUC scored 0.665 obtained using this set of parameters?

I also got the 0.6466 success rate on OTB2013 too with python2.7

Padding causes forward to produce different logits (Llama2-7b)

I found this bug too. You can test whether enable sqda or flash attention. When sqda is used, the result seems to be correct. I did not know why this...

Padding causes forward to produce different logits (Llama2-7b)

I have done some experiments. If I use the eager attention with sdpa attention mask (version==4.37.2), the results are correct. However, with the eager mode attention mask, the results are...

Padding causes forward to produce different logits (Llama2-7b)

Please take a look at the discussions above. **If left padding is used, the output of the model is wrong.** I found that the attention mask can be generated with...

Padding causes forward to produce different logits (Llama2-7b)

Thanks for your kind reply. I think I made a mistake by using different dtypes. I have fixed this by now. Sorry for the wasting of your time.