search

wvinzh / WS_DAN_PyTorch

PyTorch Implementation Of WS-DAN(See Better Before Looking Closer: Weakly Supervised Data Augmentation Network for Fine-Grained Visual Classification)
MIT License
169 stars 34 forks source link

some confusion #3

Closed boxyao closed 4 years ago

boxyao commented 4 years ago

pooling_features = raw_features*100 , Why multiply raw_features by 100?

wvinzh commented 4 years ago

Make the obtained gradient larger and speed up training，the value of raw_features is too small

boxyao commented 4 years ago

Is 100 a experience value？

boxyao commented 4 years ago

I have ran the code. The test accuracy is very well, but the train and val prec@1 and prec@5 are very samll values, such as 1.4, 1.5. Is it right？ Is there something wrong with my trainning?

wvinzh commented 4 years ago

Is 100 a experience value？

Maybe，I followed the settings in the official repo, ws_dan

wvinzh commented 4 years ago

I have ran the code. The test accuracy is very well, but the train and val prec@1 and prec@5 are very samll values, such as 1.4, 1.5. Is it right？ Is there something wrong with my trainning?

I think it is not normal, maybe you can check it carefully. 1.4 1.5 may be the loss values, not precisions