Feature request: intuitive replacement for pymc.Normal(tau=foo, observed=True, value=bar) about probability

[dustinvtran]

Have you looked at Edward2? We don't have semantics for declaring a random variable as "observed" (I think this confuses modeling/inference). But we do have a notion of binding random variables to specific values:

from tensorflow_probability import edward2 as ed

obs = ed.Normal(name="obs", loc=mean, scale=std, value=Y)

[matthew-mcateer]

Hi @dustinvtran,

I had seen that. I actually had previously used edward2, but I was attempting to use something within the larger tf.distributions or tf.contrib.distributions. I didn't know if there was something along the lies of tfd.normal_conjugates_known_scale_predictive() or tfd.normal_conjugates_known_scale_posterior() that could have acted as a substitute.

[dustinvtran]

Got it. No, it doesn't exist in TF distributions. It's Edward2's job.