About Predicted Features VS Ground Truth

[黄培松]

Hello, thank you for sharing the code。
When I learn this code, I meet a question.

As shown in the Figure, in training, it is better to use the prediction feature. But I found the code use ground truth to compute attention weight.

WaveRNN/models/tacotron.py

Line 349 in 3595219

 prenet_in = m[:, :, t - 1] if t > 0 else go_frame 

WaveRNN/models/tacotron.py

Line 240 in 3595219

 prenet_out = self.prenet(prenet_in) 

I want to know the reason why you use the Ground Truth.

[黄培松]

I got it wrong