Numerical Issues with StickBreaking in ADVI

[Dominik]

@Dominik

obj_optimizer=pm.adamax(learning_rate=1e-4, beta1=1-1e-4, beta2=1-1e-5)

I am not sure if that really solved anything or just made the divergence slower. Using it in my larger model does not seem to improve the results. It only takes much longer until the output starts to become weird.