kahnchana
student or teacher fine-tune
Dear team, Thank you again for your work on the code!
Do you fine-tune using the student model or the teacher model?
We use the teacher model. But our experiments showed that using the student does not lead to much difference.
We use the teacher model. But our experiments showed that using the student does not lead to much difference.
I don't know where the parameters in my pre training section were set incorrectly. The results I ran were only 92.01 on the ucf101 dataset and 64.17 on the hmdb51 dataset. I reproduced the weights you provided, which can reach 94.23 and 68.18, respectively
Can you disclose some information about the values of pretraining stages momentum_teacher and teacher_temp? Thank you!
@ttkxyy , I noticed that in the default setting, the code will runs for 100 epochs, which is different from the 20 epoch mentioned by papers. Did you notice the difference?
@memoiry Yes, I noticed that the parameters of SVT are the same as those of the Dino model. I think they should be modified according to the parameters proposed by the author in the paper, but I have not been able to successfully reproduce it.