Effect of min_log_var usage

[ZhuFengdaaa]

I notice that in gaussian_network.py, you use min_log_var in GaussianNetwork class to ensure the variance of z distribution not too low. However, not all subclass use this variable(e.g. GaussianBidirectionalNetwork). I am wondering why we need this variable and how this variable effect training?

Thank you :)

[jcoreyes]

The min_log_var is more of a problem for the policy (which ues GaussianNetwork) than for the decoder which uses GaussianBidirectionalNetwork. This ensures the policy gradient when training the policy does not explode.