bi-att-flow
bi-att-flow copied to clipboard
ce_loss with loss_mask however ce_loss2 without
In model.py/_build_loss()
, when computing the average loss of start position ce_loss
, loss_mask
is used to avoid counting in these samples where 0 == len(question)
. However, when computing the average loss at stop position ce_loss2
, loss_mask
is missed.