Validation accuracy is higher than training accuracy

[Sebastian-X]

Hi, thanks for your implementation!
When I run you code, I found that accuracy of "Meta-Val" dataset starts at a very high level around 0.7
, whereas accuracy of "Meta-Train" set is only 0.5. As the training goes on, this gap reduces but val accuracy is still slightly higher than train accuracy. Is this a normal phenomenon?