How to sample for both Normal distribution and Laplace distribition about mixture-density-networks-for-distribution-and-uncertainty-estimation HOT 1 CLOSED

Hi Weixsong,

I do not quite see what you mean when you are referring to my "validation code" then, if I do not answer what you need, do not hesitate to ask me again.

Following the explanation that I did in the slides and the Master Thesis, there are different ways to make predictions using MDNs models. For example, you can plot the PDF as a errorbar plot, you can use the parameters that you obtain from the MDN model to do a sampling process, you can plot a heat-map like in the end of the notebook MDN-2D-Regression.ipynb, etc. In all these cases, mu and sigma are used.

Answering your second question, when you implement an MDN model, this mixture has a combination of a certain probabilistic functions. In particular, the cases I have discussed was a mixture of all Gaussians or Laplacians functions but you could define the mixture that interested you (like here). What you should keep in mind is that you are training a neural network to be able to give you the parameters for that predefined mixture and not for any type of PDF.

from mixture-density-networks-for-distribution-and-uncertainty-estimation.