search

lukashermann / hulc

Hierarchical Universal Language Conditioned Policies
http://hulc.cs.uni-freiburg.de
MIT License
66 stars 9 forks source link

Question about Mixture Distribution #13

Closed j-sheikh closed 11 months ago

j-sheikh commented 1 year ago

Hello, first of all thank you for this great work. I have a question about the action decoder. Here you are using a mixture distribution of 10 logistic distributions. Using linear layers after the RNN you have predicted the means, logit_scales, and logit_probs for the distributions. I am wondering how exactly you train the distributions. I know that the best distribution is the most weighted, but what exactly does that mean for the other distributions in terms of mean, logit_scale, and logit_probs? What is the effect of backpropagation here? Naively, shouldn't all 10 distributions converge to the optimum?

Thank you for your time and help.

mees commented 11 months ago

Hi there, here is a detailed explanation of the mixture of logistics, hope it helps! https://github.com/Rayhane-mamah/Tacotron-2/issues/155#issuecomment-413364857