Inconsistent setting of shot 1, 2 and 3, 5, 10

[megvii-research]

Inconsistent setting of shot 1, 2 and 3, 5, 10

[bsun0802]

AFAIK, the Test Aug does not matter, you can delete that and see if the results are roughly the same.

For low shot, unfreezing all parts as in 5/10-shot incurs over-fitting and decrease the performance a bit. If you want consistency, sacrifice the performance for low shot a little. We report the best performance we can achieve here.

P.S., I personally think the current few-shot dataset are still naive and not flawless. I personally DO NOT think one can pursue an algorithm which is optimal for both 1-shot and few-shot with exact same settings.