Fixed error in loss function: log(tau(alpha)).

climax/core/loss.py

@@ -92,6 +92,7 @@ class BernoulliGammaLoss(NaNLoss):
                           torch.log(y_true + self.epsilon) -
                           y_true / (gscale + self.epsilon) -
                           gshape * torch.log(gscale + self.epsilon) -
-                          torch.lgamma(gshape + self.epsilon))
+                          torch.log(torch.lgamma(gshape + self.epsilon)
+                                    + self.epsilon))
 
         return - self.reduce(loss)