hmohebbi / valuezeroing Goto Github PK

First of all, thanks for sharing the code! I'm trying to use ValueZeroing on some experiments, and I noticed that the class BertForMaskedLM has the parameter target_index initialized as None and used in this line:

# Added by Hosein  
prediction_scores = prediction_scores[torch.arange(prediction_scores.size(0)), target_index]

But in the demo it is not passed to the model during the forward pass, so it remains None (changing the dimensions of prediciton_scores?). So, what is this parameter used for? What am I missing?

Thanks a lot in advance,
Luca