Not NAR mode in training

[youngsheen]

The training code still use the causal attention mask. You need to set full_context_alignment=True in the decoder.forward function to turn on non-causal attention mask. Is it a mistake?

[steventan0110]

Causal Mask used in reported results?

same question here. I'm benchmarking model performance with and without the causal mask and found that they result in a small change about ~1BLEU in final ASR-BLEU evaluation.
Just want to confirm with the author @Rongjiehuang if such a causal mask is used for results reported in the paper?

P.S. @youngsheen I saw you are the first author of DiffS2UT and I'm wondering if the code for that paper is released anywhere? I'm very interested in your approach as well.