Dirichlet with random shape

Questions
3

Thanks for the response.
Following your suggestion to dynamically assign zeros I have this attempt:

with pm.Model() as model:
 k = 5
 arr = []
 for j in range(k):
   arr.append(pm.Uniform('a'+str(j),lower=0,upper=0.99))
 w = pm.Dirichlet('w', a=tt.stack([arr[i] for i in range(k)]), shape=k)
 du = pm.DiscreteUniform('du',lower=0, upper=1,shape=k)
 mu = pm.Normal('mu', 0., 10., shape=k)
 tau = pm.Gamma('tau', 1., 1., shape=k)
 x_obs = pm.NormalMixture('x_obs', w, np.multiply(mu,du), tau=np.multiply(tau,du), observed=x)

Basically what I’m doing is to multiply by 0 or 1 the elements of w, mu, and tau. the 0's and 1's are those of the DiscreteUniform

It’s not working. The error is (only showing last fragment):

...........................................................................
/usr/local/lib/python2.7/dist-packages/pymc3/step_methods/hmc/base_hmc.py in astep(self=<pymc3.step_methods.hmc.nuts.NUTS object>, q0=array([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
   0., 0.]))
144                 self.iter_count,
145                 None,
146                 None,
147             )
148             self._warnings.append(warning)
--> 149             raise SamplingError("Bad initial energy")
150 
151         adapt_step = self.tune and self.adapt_step_size
152         step_size = self.step_adapt.current(adapt_step)
153         self.step_size = step_size

SamplingError: Bad initial energy

I think the DiscreteUniform in particular lower=0 is generating the error.

If I change the bounds to something bigger than 0 it doesn’t throw the error but after sampling all du's are 1, defeating the purpose.

How can I implement the idea correctly?
Thanks