Fix 2948

[junpenglao]

Reverse the logpt tensor computation back to pre #2499.
close #2948

[kyleabeauchamp]

Maybe we should add in a test from #2948? E.g., ideally we want to end up in a place where tests guarantee that both #2948 and #2482 are resolved.

[junpenglao]

This will not effect the fix for #2482.
I am still trying to find out why it only fail for Lognormal... if it really is an edge case I will then add a test accordingly.

[junpenglao]

So far this is really weird...

ndim = 3
with pm.Model() as model:
    sigma = pm.Lognormal('sigma', 
                         mu=np.zeros(ndim),
                         tau=np.ones(ndim), 
                         shape=ndim)  # variance for the correlation matrix
    nu = pm.HalfCauchy('nu', beta=10)

factors = [var.logpt for var in model.basic_RVs] + model.potentials
logp = tt.sum([tt.sum(logpt) for logpt in factors]) --> fail
logp = tt.add(*map(tt.sum, factors)) --> works
func = pm.model.ValueGradFunction(logp, model.basic_RVs)
func.set_extra_values(model.test_point)
q = np.array([2.30258509, 1., 1., 1.])
func(q)

And the error only appears using Lognormal when the tau or sd is multi dimensional...

[junpenglao]

I will merge this if no more comment.

[ColCarroll]

LGTM, though the previous code also looked good. Glad you added some tests!

[ColCarroll]

Possible performance hit from this change? http://pandas.pydata.org/speed/pymc3/#benchmarks.NUTSInitSuite.track_glm_hierarchical_ess

[junpenglao]

Thanks for the heads up. I am a bit surprised that it effected the effective sample size - will report back.