Using a bounded variable within a Mixture

Questions
1

I am trying to use a Bounded variable within a Mixture model. I have the following so far:


with pm.Model() as model:
    alpha = pm.Gamma('alpha', 1,1)
    beta = pm.Beta('beta', 1, alpha, shape=K)
    w = pm.Deterministic('w', stick_breaking(beta))
    #a = pm.Uniform('a', 0, 50, shape=K)
    #b = pm.Uniform('b', 0, 50, shape=K)
    mu = pm.Uniform('mu', 0, 50, shape=K)
    bp = pm.Bound(pm.Poisson, lower=1)('bp', mu=mu)
    x = pm.Mixture('obs', w, bp.dist(), observed=df['x1'].values)

It throws the following error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-121-ee34f0f15713> in <module>()
      8     mu = pm.Uniform('mu', 0, 50, shape=K)
      9     bp = pm.Bound(pm.Poisson, lower=1)('bp', mu=mu)
---> 10     x = pm.Mixture('obs', w, bp.dist(), observed=df['x1'].values)

AttributeError: 'FreeRV' object has no attribute 'dist'

How do I correct this?

2

I think you should write bp as bp = pm.Bound(pm.Poisson, lower=1).dist(mu=mu)

3

@aloctavodia
You mean like this?


with pm.Model() as model:
    alpha = pm.Gamma('alpha', 1,1)
    beta = pm.Beta('beta', 1, alpha, shape=K)
    w = pm.Deterministic('w', stick_breaking(beta))
    #a = pm.Uniform('a', 0, 50, shape=K)
    #b = pm.Uniform('b', 0, 50, shape=K)
    mu = pm.Uniform('mu', 0, 50, shape=K)
    bp = pm.Bound(pm.Poisson, lower=1).dist( mu=mu)
    #x = pm.Mixture('obs', w, pm.Poisson.dist(mu), observed=df['x1'].values)
    x = pm.Mixture('obs', w, bp, observed=df['x1'].values)

I get the following error:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
C:\Anaconda3\envs\stpy35\lib\site-packages\pymc3\distributions\mixture.py in _comp_modes(self)
     95         try:
---> 96             return tt.as_tensor_variable(self.comp_dists.mode)
     97         except AttributeError:

AttributeError: '_DiscreteBounded' object has no attribute 'mode'

During handling of the above exception, another exception occurred:

TypeError                                 Traceback (most recent call last)
<ipython-input-15-811a1e7927a9> in <module>()
      9     bp = pm.Bound(pm.Poisson, lower=1).dist( mu=mu)
     10     #x = pm.Mixture('obs', w, pm.Poisson.dist(mu), observed=df['x1'].values)
---> 11     x = pm.Mixture('obs', w, bp, observed=df['x1'].values)

C:\Anaconda3\envs\stpy35\lib\site-packages\pymc3\distributions\distribution.py in __new__(cls, name, *args, **kwargs)
     34                 raise TypeError("observed needs to be data but got: {}".format(type(data)))
     35             total_size = kwargs.pop('total_size', None)
---> 36             dist = cls.dist(*args, **kwargs)
     37             return model.Var(name, dist, data, total_size)
     38         else:

C:\Anaconda3\envs\stpy35\lib\site-packages\pymc3\distributions\distribution.py in dist(cls, *args, **kwargs)
     45     def dist(cls, *args, **kwargs):
     46         dist = object.__new__(cls)
---> 47         dist.__init__(*args, **kwargs)
     48         return dist
     49 

C:\Anaconda3\envs\stpy35\lib\site-packages\pymc3\distributions\mixture.py in __init__(self, w, comp_dists, *args, **kwargs)
     62 
     63         try:
---> 64             comp_modes = self._comp_modes()
     65             comp_mode_logps = self.logp(comp_modes)
     66             self.mode = comp_modes[tt.argmax(w * comp_mode_logps, axis=-1)]

C:\Anaconda3\envs\stpy35\lib\site-packages\pymc3\distributions\mixture.py in _comp_modes(self)
     96             return tt.as_tensor_variable(self.comp_dists.mode)
     97         except AttributeError:
---> 98             return tt.stack([comp_dist.mode for comp_dist in self.comp_dists],
     99                             axis=1)
    100 

TypeError: '_DiscreteBounded' object is not iterable
4

ups!

I guess this will also fail, right?

bp = pm.Bound(pm.Poisson, lower=1).dist(mu=mu, shape=K)

so, try this

bp = [pm.Bound(pm.Poisson, lower=1).dist(mu=mu[i]) for i in range(K)]
1 Like
5

@aloctavodia
Yes, that second one works. Thank you very much.

However, how do I refer to the variable by name? So I can grab it from trace, like so trace[‘myvar’]. Here is the code in case someone else needs to use similar tricks:

with pm.Model() as model:
    alpha = pm.Gamma('alpha', 1,1)
    beta = pm.Beta('beta', 1, alpha, shape=K)
    w = pm.Deterministic('w', stick_breaking(beta))
    #a = pm.Uniform('a', 0, 50, shape=K)
    #b = pm.Uniform('b', 0, 50, shape=K)
    mu = pm.Uniform('mu', 0, 50, shape=K)
    bp = [pm.Bound(pm.Poisson, lower=1).dist( mu=mu[i]) for i in range(K)] 
    #x = pm.Mixture('obs', w, pm.Poisson.dist(mu), observed=df['x1'].values)
    x = pm.Mixture('obs', w, bp, observed=df['x1'].values)
6

Not sure if that’s possible with the current implementation of Mixture. Maybe @AustinRochford has a solution.

7

ok, thank you @aloctavodia. Is it possible for me to message you in private, so perhaps you could show me around the codebase and that way help myself quicker and perhaps other folks as well?

8

You can contact me in private, no problem. But if you have a question that may also help others, is better to ask it publicly so other can benefit from it and also you may get a better/faster answer from others than from me.

1 Like
9

I think bp needs to have shape > 1 in order for a mixture to make sense.