Sampling Strategy

[sunprinceS]

In MAML, FOMAML algorithm, we split each task into D_train, D_test, and use D_test as the last step in each run_task.

But in current implementation of FOMAML, the sampling strategy follows the one in Reptile, won't split D_train and D_test, but random draw.

So we might add dome different sampling strategy in meta_interface

1. meta-split: split train into inner-train and inner-test
2. meta-split-dev: use train as inner-train, dev as inner-test (But in that case, we should make multi_interface also train on dev, too).

[sunprinceS]

HYL thinks that the original setting is fine