I am not familiar with the ddpm, but when try to train the ddpm in of tortoise-tts in this project, there are some different loss components, "loss_ddpm_total", "loss_diffusion_loss", "loss_mel_surrogate", "loss_upper_mse", "loss_var_loss", can anyone help to explain what are those components do? and how can I judge the diffusion model converage with any of the above components? and , If we do fine-tune training with AR model output latent, is there any change about the loss usage above?
I am not familiar with the ddpm, but when try to train the ddpm in of tortoise-tts in this project, there are some different loss components, "loss_ddpm_total", "loss_diffusion_loss", "loss_mel_surrogate", "loss_upper_mse", "loss_var_loss", can anyone help to explain what are those components do? and how can I judge the diffusion model converage with any of the above components? and , If we do fine-tune training with AR model output latent, is there any change about the loss usage above?