Training using 'protonet', 'matchingnet', 'relationnet', 'relationnet_softmax', 'maml', 'maml_approx' methods with an unbalanced dataset.

[adnan119]

Thanks for this amazing work @jbhuang0604 @ycliu93 @wyharveychen

Currently, I'm trying to train on FIGR-8 dataset using the models and methods available in this repo.

Training using the baseline and baseline++ works great, However, the issue arises when trying to train with protonet and other above mention methods which uses n_shot, n_way, n_support as parameters for defining the batch_size and other model and training parameters. Turns out that FIGR-8 is an unbalanced dataset with each class containing at least 8 images.

What I'd like to know is if there's any way to overcome/bypass this issue without modifying the original dataset and train using the above-mentioned method.