SeesawLoss_pytorch

This implementation is based on bamps53/SeesawLoss. His implementation only involves mitigation factor, no compensation factor.Following his implementation, i added compensation factor to loss.

useage

from seesawloss import DistibutionAgnosticSeesawLossWithLogits
num_labels = 10
loss_fn = DistibutionAgnosticSeesawLossWithLogits(num_labels=num_labels)
loss = loss_fn(logits, label)

preds: logits

label: not one-hot label

If there is any problem with my implementation, please let me know. thanks!

Citation

1. This is unofficial pytorch implementation for SeesawLoss, which was proposed by Jiaqi Wang et. al. in their technical report for LVIS workshop at ECCV 2020.