Hi, I find that the label_pad_token id is different between layoutxlm and ernie layout in token classification. In layoutxlm, the label_pad_token id is -100 and will be ignored by pytorch loss functions url. However the label_pad_token id is 0 in ernie layout and will be calculated loss. url2
Hi, I find that the label_pad_token id is different between layoutxlm and ernie layout in token classification. In layoutxlm, the label_pad_token id is -100 and will be ignored by pytorch loss functions url. However the label_pad_token id is 0 in ernie layout and will be calculated loss. url2