search

ManLiuCoder / PSVMA

18 stars 4 forks source link

accuracy for AWA2 much worse than reported #8

Open RowenaHe opened 5 months ago

RowenaHe commented 5 months ago

Hi! I ran your code on my local machine. The accuracy of CUB and SUN is about the same as the ones reported in the paper. But for the AWA2 dataset, the reproduce result only 69% for H and 67% for T. I did NOT change any hyperparameter. Can you please check about this problem?

image

ram-neuralcoder commented 4 months ago

Hi, I ran this code, but why don't the model's results converge? I didn't change the code elsewhere except for loading the model. The loaded vit is checked to be trainable.