Closed chagmgang closed 10 months ago
Thank you for reply. Is it splited randomly?
Hi, sorry for missing previous comment.
Many of the classification dataset do not provide official train/val/test split, therefore we have to do this by our own. In our implementation, for each class, we use the first 80% samples (starting from something like 001.jpg
, 002.jpg
...) for training, and use the remaining for testing.
Thank you reply.
Hi, sorry for missing this detail in our paper.
In classification experiments of both linear-prob and k-NN, we use a spliting ratio of train:test=8:2.