Currently when we read the validation data file during training, we insert INNER to represent the inner nodes. This is the reference output sequence. But the model will be in eval mode and so it will not output these INNER tags. So we need a separate TGT tag for reading validation data.
Currently when we read the validation data file during training, we insert INNER to represent the inner nodes. This is the reference output sequence. But the model will be in eval mode and so it will not output these INNER tags. So we need a separate TGT tag for reading validation data.