yanggeng1995 / vae_tacotron

MIT License
51 stars 19 forks source link

What reference does vae_weight follow? #9

Open sungjae-cho opened 4 years ago

sungjae-cho commented 4 years ago

Hello, @yanggeng1995
You implemented vae_weight as such. What reference does vae_weight follow?

Thank you in advance!

yanggeng1995 commented 4 years ago

Hello, vae_weight is only used for training.

Sungjae Cho notifications@github.com 于2020年4月12日周日 下午10:49写道:

Hello, @yanggeng1995 https://github.com/yanggeng1995 You implemented vae_weight as such https://github.com/yanggeng1995/vae_tacotron/blob/df7d02c92d28ed16b65ea42d27735e49882bf978/util/util.py#L24. What reference does vae_weight follow?

Thank you in advance!

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/yanggeng1995/vae_tacotron/issues/9, or unsubscribe https://github.com/notifications/unsubscribe-auth/AD73U3VTBYZAUD7RM67BPPTRMHIGNANCNFSM4MGOB2SA .

sungjae-cho commented 4 years ago

Hello, what I am curious about is what paper or reference did you follow when you implemented vae_weight. I want to know what weight scheduling method vae_weight follows.

yanggeng1995 commented 4 years ago

Hello, what I am curious about is what paper or reference did you follow when you implemented vae_weight. I want to know what weight scheduling method vae_weight follows.

Hello. The paper does not clearly explain how vae_weight is adjusted according to step, only that simulated annealing is used. This is just what I realized according to my personal ideas.

sungjae-cho commented 4 years ago

Simulated annealing (SM)!

Could you answer the following questions? @yanggeng1995

  1. Does your SM work well with training GMVAE-Tacotron?
  2. I've first seen this kind of SM, which enforces the VAE weight once during the period and gradually increases the weight. Could you tell me where you get inspiration this kind of SM?