Greetings,
I am trying apply knowledge distillation for domain adaptation problem in NMT, I understood that to create my custom loss function I need to subclass the Trainer class and override the compute_loss function. As I’m following Sequence level distillation, I am required to pass data from 6 different domains, calculate the individual loss per domain then calculate the global loss. It will wonderful if someone could point me to any resource which shows passing multiple datasets into the Trainer Class