the results of evaluation online and evaluation alone are different

[windyjunfeng]

Hi, I met a tough problem. The results of evaluation after training an epoch were largely different from the ones I evaluated by loading the saved weights on the same dataset. Maybe it is because the difference of ImageDataset. I only used EnsureChannelFirst() transform and found the random seed in __getitem__ function of ImageDataset codes. How can I solve it? I want to get the same results whether online or alone. Thank you very much.