Error while computing GMM Log Loss when predict_length = 5 about livespeechportraits HOT 2 CLOSED

'predict_length' is a decrypted option in my history experiments, which I tried to predict more than one frame's results but I found it useless finally. I always output ONE frame results for each forward pass.
Sorry that I didn't carefully check the training-related codes and clean the useless parts.

GMMLogLoss is a generalized GMM loss, which means you can model the target using several Gaussian distributions. In that case, you need to output each probability for each gaussian and choose one gaussian as your distribution at the time. In my case, I only use one gaussian to model the head pose, and therefore the probability for each gaussian is meaningless. That's why you see 25 values (24 + 1). I output this probability to keep the completeness of the GMM loss (but it is useless actually).

Hope the above helps.

from livespeechportraits.

Oh I see!
It was quite hard to keep track of all the config parameters throughout the code..
Makes sense now, thanks for the help! And nice work 👍

from livespeechportraits.