I notice that the default value of the aux loss multiplier is very large, i.e. 46 in the code (line396).
If my understanding is correct, does that mean during learning, the method put very large attention on learning from the supervised part, let’s say 46 on the supervised part and 1 on the unsupervised part?
Here is the tutorial of the semi-supervised VAE. Corresponding to the objective function in section Third Variant: Adding a Term to the Objective of the tutorial, can I consider the alpha equals to 1 and the alpha’ equals to 46 here?