Validation accuracy/perplexity conditions on "gold" target tokens?

Hi,
In OpenNMT-py/onmt/Trainer.py does the validation condition of true target tokens?
My understanding is that the validation scenario should replicate the test scenario as much as possible and therefore validation decoding should not condition on the true/gold target side tokens.

Is there a way to ensure this in OpenNMT-py?

1 Like