search

rishikksh20 / vae_tacotron2

VAE Tacotron 2, an alternative of GST Tacotron
MIT License
87 stars 29 forks source link

What a successful error curve should look like？ #8

Closed xiaoyangnihao closed 4 years ago

xiaoyangnihao commented 4 years ago

@rishikksh20 Hi, I use this repo with Blizzard2013 dataset instead of ljspeech dataset with default settings. I want to know whether I am training this vae_model right. or not ? What a successful error curve should look like with this repo? The followings are my curves. I wonder whether the difference between kl_loss and reconstruction_loss is too wide. Thanks for any help?

image image image

rishikksh20 commented 4 years ago

15k steps from where VAE loss kicked in that's why kl loss increase bit.

rishikksh20 commented 4 years ago

After loss curve seems ok but stats and kl loss are not converging, for understanding those curve you need to understand the Variational autoencoder training process.. I am not well remembered how should curve be like in I had trained this.

xiaoyangnihao commented 4 years ago

15k steps from where VAE loss kicked in that's why kl loss increase bit.

Isn't vae loss increased by kl_loss，why do you think it is the opposite, vae_loss ==> kl_loss ?

xiaoyangnihao commented 4 years ago

After loss curve seems ok but stats and kl loss are not converging, for understanding those curve you need to understand the Variational autoencoder training process.. I am not well remembered how should curve be like in I had trained this.

thanks, I will find it out. By the way, I want to make sure that whether the init_vae_weights=0.001 and vae_weight_multiler=0.002 is two big. The mel_loss and kl_loss are not on the same level of magnitude. mel_Loss maybe 0.01, but the kl_loss maybe 300 * 0.3 or 0.