a question of acc

[Amos1109]

Hello, thank you very much for your work! But when I was performing the experiment (office_home, C->A, 3shot), when the number of iterations exceeds 9000 times, acc, acc1, mAcc, mAcc1 do not change, always 69.80 (acc1, acc2, 71.40 in paper) and 64.11(mAcc1, mAcc2), am I missing an important step?

[Chief Accelerator]

Hi, as the scale of office-home is very small, the accuracy may fluctuate a lot for different runnings. Besides, we empirically found that MCL may overfit the office-home if you train it in a very long time, so we report the best-accuracy for office-home, not the final accuracy.