I am trying to implement Dynamic Samplers in Pytorch. This seems impossible with the existing train_loader api's. As the weighted the samplers only allow weights to be set once. I was thinking that weights could change based on the loss value for each and individual data point. This can be done without train_loader API but I think this a feature that could be included in a future release.

Something as simple as Loss Based sampling can't be done right now.

cc @ssnl @VitalyFedyunin @ejguan