about dataset spiltting

[raymond00000]

Hi,

Sorry I am confused on the data splitting steps and codes.
Please kindly correct me if I got it wrong.

Omniglot has 50 classes. it allocated 30 classes in its original train split, and 20 in test split. these 20 is not used during model-training.

Assume we do 10 classes 5 shot classification in MAML.
in inner loop, we will randomly picked 10 classes from that 30 classes. For each class, we sample 5 pictures.
so, we will have 50 samples as support set, then we use these samples to update the base-model, so base-model becomes meta-model.
then, we will retrieve some samples (how and how many?? I am very confused on this sampling step) as query set, then we calculate loss with meta-model using these samples but to update the base-model.

my question is about the query set.

1. At the very beginning, do I have to split the 30 classes into 2 groups, say, 15 for support set, 15 for query set? I think I don't have too.
2. within the same class, does it have to be non-overlapping between the query set and support set? In another words, for one specific sample, will it be in support set in one mini-batch and in query set in next mini-batch? I think it is yes.

Many thanks for answering.