Inspiring work. After reading the paper and guidelines, I have some questions about the remask module. In the inference process, low-confidence remasking, in my understanding, is a heuristic adjustment of the quality of the non-autoregressive generated sequence, which is necessary, otherwise the generation effect will only be worse. I am curious whether remasking will increase or decrease the difficulty of the next step of denoising in the model, resulting in insufficient or overflow in the final inference steps. How is this guaranteed and evaluated?
AI Analysis
This issue appears to be discussing a feature request or bug report related to the repository. Based on the content, it seems to be resolved. The issue was opened by bansky-cl and has received 1 comments.
Add a comment
Comment form would go here