All the term of training losses are 0.000 except mel loss

yl4579 / StyleTTS2

StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models

MIT License

4.95k stars 417 forks source link

All the term of training losses are 0.000 except mel loss #159

Closed jirenhui closed 11 months ago

jirenhui commented 11 months ago

Great work! However, When I reused the LJSpeech to train the model, I encountered some problems as shown in the figure below. train_list, val_list and OOD_text are all original in the project. Can you help me solve it?

martinambrus commented 11 months ago

That's normal for the first 49 epochs on 1st training stage. Once you get to epoch 50, you'll see the other losses as well.