Hey,
I would like to build a DP with components using a LKJ-Prior. Since I know that there is no shape-argument for the LKJ-distribution in pymc3, I have tried to use tt.stack to generate my components. But I don’ t really get satisfying results…
So I was wondering if someone has other ideas on how to bypass this problem.
So far my approach was this:
packed_L = tt.stack([pm.LKJCholeskyCov('packed_L_{}'.format(k), n=1, eta=2.,
sd_dist=pm.HalfCauchy.dist(beta=2.5)) for k in range(K)])
Thanks in advance
Leon