Batch running statistics being updated at validation time

[AntreasAntoniou]

In line 340 of the maml.py file you use:

pred = self.net(support_x[i])

Which makes training to default to True. Do we really want the batch running stats to be updated at inference time?

[dragen1860]

Hi, About this problem, setting it to True or False does not have too much difference according to batchnorm setting, if i not wrong.
Especially for meta-learning problem, the meta-train includes 2 sub stages: training and test, so is meta-test.
Hence, I found it obscure to set the flag to True or False for meta-learning problem.
You can try to set it to False for meta-test, and see any improvement。