Use sigmoid activation for p; optimize logarithms of scale, shape.

climax/core/loss.py

@@ -70,11 +70,11 @@ class BernoulliGammaLoss(NaNLoss):
         # parameters: ensure numerical stability
 
         # clip probabilities to (0, 1)
-        p_pred = F.sigmoid(y_pred[:, 0, ...])
+        p_pred = torch.sigmoid(y_pred[:, 0, ...])
 
         # clip shape and scale to (0, +infinity)
-        gshape = torch.exp(y_pred[:, 1, ...].clamp(min=self.epsilon))
-        gscale = torch.exp(y_pred[:, 2, ...].clamp(min=self.epsilon))
+        gshape = torch.exp(y_pred[:, 1, ...])
+        gscale = torch.exp(y_pred[:, 2, ...])
 
         # negative log-likelihood function of Bernoulli-Gamma distribution