AprilYapingZhang/Seq2SeqAdapt

No improvement in accuracy during training

XueyingDing123 opened this issue · 1 comments

Hello, I found that the recognition accuracy has not been improved when I trained according to the readme. I am wondering if friends with the same problems or author could provide some suggestions. Thank you very much!

The recognition accuracy rises in fluctuation during training process, due to the adversarial traning between feature domain adaptation and decoding loss. You can observe the improvement of performance after 10000 training steps.