Dear all,
i use the GraphConvModel with uncertainty and was interested in the “epistemic” and “aleatoric” contribution to the total uncertainty. I found that the epistemic contribution is basically zero for all molecules (not only for the delaney data) and for dropout ratios = 0.1, 0.5, 0.9. I’m not an expert, but after looking at the code, I wonder whether dropout is actually turned on during prediction.
Best,
Alex
