CategoricalGibbsMetropolis samples differently when specified explicitly

Questions
1

Hi. When a categorical variable is sampler is assigned automatically as CategoricalGibbsMetropolis it seems to behave differently to when specified as a keyword argument.
E.g.

n = 10
X = np.random.rand(n)
K = 1+np.arange(n)*0.1
Y = K[3]*X + np.random.randn(n)*0.01

with pm.Model() as model:
    ind = pm.Categorical('ind', p=np.ones(n)/n)
    k = t.shared(K)
    mu = k[ind]*X  
    y = pm.Normal('y', mu=mu, sd=0.01, observed=Y,shape=n)
    trace = pm.sample(10000) 
    outind = trace.get_values(ind)

will return the correctly sampled posterior in the trace.
Whereas specifying the same sampler explicitly with:

step = pm.CategoricalGibbsMetropolis([ind]) 
trace = pm.sample(10000, step=[step])

the trace instead returns the prior distribution.
Why is the prior distribution being returned here? Or am I misunderstanding the step keyword?

2

Hmm, I cannot reproduce the problem, could you please try update to pymc3 master?

3

Yep, that sorted it out thanks.