class imbalance handling w/ResNet strikes back

[kyle-dorman]

Hi,

I have read through the ResNet Strikes Back paper and taken a look at your code. I am trying to train a model (pretrained or from scratch) on a dataset with ~1500 classes and ~200000 images with class imbalance (min samples 10, max ~1K).

I was just curious if you experimented at all with setting the pos_weight parameter when training "Resnet strikes back" models using the BCE loss? My intuition was to make use of it in my problem but I didn't see you experiment with it at all when training and I know imagenet also has a class imbalance. If not, did you take any other actions to handle the class imbalance? (I don't see any called out in the paper or code, but figured I would ask).