In order to promote this question, till now I’ve found that when I’m performing next command:
o = Mixture(cat=cat, components=components, value=tf.zeros_like(o_ph))
, the edward infrastructure expects batch shape of cat and components be the same.
While cat has scalar dimension:
(none is number of sample points)
, the components in my case is Normal distribution and its batch_shape is:
(2 is current dimension of my output)
When creating Mixture, I’m getting error:
ValueError: Shapes (?,) and (?, 2) are not compatible
So, why does Mixture expect cat and component to share dimensions?
And how can I fix it?