You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi, thanks for pointing out this issue. In my original code, I use the early released dataset (before 31 Aug 2017), which has overlap images between train and test set, so the result is slightly (+0.4% accuracy) higher. The latest released dataset has fixed this problem, so if you use the latest dataset, you will achieve a slightly lower performance (both baseline and baseline+RE). More details please refer to zalandoresearch/fashion-mnist#44
However, our method can also improve the performance on the latest release dataset.
I trained a model for Fashion-MNIST. Unfortunately the performance is not as good as the result reported in the paper.
The best accuracy is 95.8% to me, while the error is 3.65 in the paper when training WRN-28-10 for Fashion-MNIST.
I wonder if the authors can provide the command line argument that used in the paper, so that people (like me) can reproduce the experiment.
The text was updated successfully, but these errors were encountered: