[TTS]Fix diffusion wavenet denoiser final conv init param by HighCWu · Pull Request #2868 · PaddlePaddle/PaddleSpeech (original) (raw)

PR types

Bug fixes

PR changes

Models

Describe

When I used WaveNetDenoiser to train diffusion-related tasks, I found that the loss did not decrease. After compared with diffsinger again, I modified the initialization of the last convolution. The meaning of stack and dilation_by_cycle_length in diffsinger is different. stack is actually equal to the number of layers divided by dilation_by_cycle_length.

我自己使用WaveNetDenoiser训练diffusion相关任务时发现损失不下降，与diffsinger再次对齐后修改了最后一个卷积的初始化，并且stack与diffsinger中的dilation_by_cycle_length含义不同，stack实际上等于layer数除以dilation_by_cycle_length。