When creating a custom loss for a model using intermediate tensorflow probability DistributionLambda layer, it was easy without Eager execution, as the layer could
DistributionLambda