how to train beta in translated sigmoid function

[wdnmd80]

In Methods---construction of entropy network: "beta is an overall scaling (trainable, constrained to the positive domain)"

but in your script beta is not trained but give it to vae model directly, can you give me a reason why you set uncertainty_temperature equal to -3.5 in distances.ipynb

or can you suggest a loss function to optimize this parameter?