We clip gradients when their global norm ex-ceeds 1 and use parallel-mode monotonic attention with initial en-ergy function scalar bias set to -1.
https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/contrib/seq2seq/monotonic_attention
https://arxiv.org/pdf/1704.00784.pdf
标贝数据集文本的处理脚本: /home/hujk17/data_BZNSYP/ProsodyLabeling
LJ处理脚本的数据:
tacotron/preprocessor.py
其中删掉了很多. 13000=>7980 要改正.
mix他们http://219.223.184.100:8880/notebooks/Mix_phoneme_G2P_demo.ipynb