KevinMIN95 / StyleSpeech

Official implementation of Meta-StyleSpeech and StyleSpeech
MIT License
241 stars 39 forks source link

Pre-trained Style Encoder gives Nan values #24

Open Conscht opened 5 months ago

Conscht commented 5 months ago

Issue: Using the extracted pre-trained style encoder results in output embeddings with NaN values, but only during training.

Resolution: Training was initially running on mixed precision at 16 bits. However, changing it to 32 bits resolved the issue, and only correct embeddings were generated.