Some discrepancies from the paper

[ericjang]

For the Mini-ImageNet ResNet-based encoder, I believe the authors use mean pooling instead of max pooling, as done by this implementation. I am currently unable to reproduce the mini-imagenet results with your codebase (even train accuracy does not go up), and am hunting for discrepancies :)

[eambutu]

Awesome thanks so much! I hadn't achieved parity with MiniImagenet yet when I'd stopped working on this, so I appreciate the help! I finally have access to compute resources again, so I might try again this weekend.

One thing that might be worth a shot is trying loading/processing MiniImagenet in a different way: in other meta-learning repos using Miniimagenet, they download the images from the link in (dragen1860/LearningToCompare-Pytorch#4) and do some other normalization (see: https://github.com/dragen1860/MAML-Pytorch/blob/master/MiniImagenet.py#L55).

[suoql]

Hey guys, thanks for sharing the suggestions! I added image transformation based on your dataloader. But the train accuracy still cannot go up. Are you guys able to reproduce the results on miniImagenet so far?