Closed DreamtaleCore closed 2 years ago
Training the audio2headpose module using smooth loss is one of my history experiments. Actually, I didn't use this smooth loss finally, where I only use the probabilistic loss as demonstrated in the paper. I think this smooth loss doesn't work obviously as I remember, so it is deprecated.
Sorry that I didn't clean up the training-related codes clearly and they confuse your training.
Gaussian Mixture Model is just a multi-gaussian version loss, and I just use one gaussian so it degrades to a single Gaussian distribution. I describe it in the code comments on in GMM loss function.
I didn't know any alternative to this loss and this is my implementation. You can check it on the internet and write your own version to speed it up.
Thanks!
Hi, I'm trying to repeat the training part of
audio2headpose
these days. I have two questions about the implementation.mu_gen=Sample_GMM ...
(Line-103) inaudio2headpose_model
benefit to the performance? Besides, I have found 'We also tried with a Gaussian Mixture Model but found no obvious improvement' in the paper, but I am a little confused. Are those the same thing? It seems the implementation of Eq(8) is theSample_GMM
function (please correct me if I am wrong).Sample_GMM
is rather low. When using it (setsmooth_loss > 0
), it needs ~2h for one epoch. I find that there are too many for-loops (line-99) and CPU operation. Are there other alternatives?