Hey all! I was wondering if it’s safe code or good practice to directly call distribution.logits if I want to obtain the log probabilities for the action probabilities such that I can do entropy calculations/backprop. Or should I directly calculate them myself (with a torch.log + epsilon)?