Regarding the issue of pre-trained weights. #21
Description
Hello, I am amazed by your outstanding work. However, I have a small question that puzzles me. Based on your code's parameter settings, the accuracy only reaches the second-best accuracy shown in the green box, but it cannot reach the optimal accuracy indicated by the red line. I am not sure whether the RN101-Superglobal method corresponds to a different pre-trained weight compared to RN101-CVNet, but as far as I know, the publicly available pre-trained weights only include the CVNet version. Has anyone else encountered the same issue?
