Handling imbalanced classes (`crit` doesn't exist anymore?)

[Soo (Hyeonsoo Lee)]

It’d be nice to check there is an oversampling callback implemented by awesome person!
Usually people say oversampling or undersampling methods are better than changing loss weight.
It’s just implemented a week ago

Oversampling Callback

Hello all, I wrote my first callback, meant for performing oversampling in fastai when training on imbalanced datasets. I started using oversampling on a severely imbalanced medical dataset, and preliminary results seem to show that oversampling significantly helps training. Here is the Callback: from torch.utils.data.sampler import WeightedRandomSampler class OverSamplingCallback(Callback): def __init__(self): self.labels = learn.data.train_dl.dataset.y.items _, counts = …