SNRE-A logits to likelihood conversion

[jsv1206]

In the following code for SNRE-A loss function, I noticed that logits to likelihood conversion is done using sigmoid and not softmax. Is there any reason for not using softmax?

sbi/sbi/inference/snre/snre_a.py

Line 117 in fe55b1c

likelihood = torch.sigmoid(logits).squeeze()