Something wrong in the config of the last max pooling layer.

[MrDavidG]

The setting of the last pooling layer is ('max_pool2d', [2, 1, 0]), which I think should be [2, 2, 0]. Maybe this is the reason of the poor performance?

[IandRover]

From the original work, it seems that the "same" padding is used instead of the "valid" padding. Therefore the padding in the config in miniimagenet_train.py should change from 0 to 1, i.e. '('conv2d', [32, 3, 3, 3, 1, 0])' to ('conv2d', [32, 3, 3, 3, 1, 1]),
After changing the padding in the config from 0 to 1, one can change the config of ('max_pool2d', [2, 1, 0]) to ('max_pool2d', [2, 2, 0]).