The problem about the beginning index

[Loose-Gu]

The scorer computes the scores by sequence_cross_entropy_with_logits(). I notice that the begining index of the para is different from the implementation in EPR.
in UDR:
loss_list = sequence_cross_entropy_with_logits(logits=output.logits[:, :-1].contiguous(), targets=entry.input_ids[:, 1:].contiguous(), weights=pad_mask, average=None)
in EPR:
loss_list = sequence_cross_entropy_with_logits(logits=output.logits, targets=entry.input_ids[:,1:], weights=pad_mask, average=None)
So I wander what actually the input is and find this in scorer_dsr.py
tokenized_example = self.tokenizer.encode_plus(enc_text, truncation=True, add_special_tokens=False, return_tensors='pt') tokenized_labels = self.tokenizer.encode_plus(test_answer, truncation=True, add_special_tokens=False, return_tensors='pt')
Since the special tokens aren't add into the inputs, Why do we need to exclude the first of the inputs and the end of the logits?

[LeeSureman]

In UDR's implementation, we score examples batch by batch, when EPR's code scores examples one by one. To implement this batch-parallel example scoring, we adjust the example's pad position for tokenization, which leads to the mentioned difference.