BVAE seems to learn more on big trees

[juripetersen]

One approach to get rid of this problem could be log scaling in the loss function in order to make every loss term as important as the other.