Manually sampling from prior predictive

Questions
1

I would like to sample from prior predictive manually. I’m referring to the documentation in the section Using PyMC Distributions without a Model of this tutorial, but I get an error:

y = pm.Binomial.dist(n=10, p=0.5)
y.random(size=3)

AttributeError: The rv.random() method was removed. Instead use rv.eval().`

I’m using 4.0.1 and see that the doc is for version 3, so is there a different syntax for 4?

2

The updated documentation is here.

3

I think this is what you are looking for: pymc.draw — PyMC dev documentation

4

The updated doc is the same as for version 3. With the pm.draw syntax, how would I sample from prior predictive when my likelihood is a mixture model?

with pm.Model() as model_mix:
    mu1 = pm.Normal('mu1', mu = 1.73, sigma = 0.035)
    sigma1 = pm.HalfNormal('sigma1', .03)
    mu3 = pm.Normal('mu3', mu=1.85, sigma = 0.025)
    sigma3 = pm.HalfNormal('sigma3', .02)

    norm1 = pm.Normal.dist(mu=mu1, sigma = sigma1)
    norm3 = pm.Normal.dist(mu=mu3, sigma = sigma3)

    w = pm.Dirichlet('w', a=np.array([1, 1]))

    like = pm.Mixture('like', w=w, comp_dists = [norm1, norm3], observed=obs_array)

    trace_mix = pm.sample()
5

Have you seen pymc.sample_prior_predictive — PyMC dev documentation ?

6

I’ve checked it out, but it’s a little too complex . Was hoping there would be a simpler way to recreate samplie_prior_predictive by sampling the parameters separately.

7

You can call pm.draw on a likelihood (or any other model variables)

8

So pm.draw works well to sample either from the separate priors (mu1, norm1, sigma1, sigma3) or from the likelihood. So that’s good. However, how would I sample from the full prior distribution (basically specifying the mixture model but without conditioning the data on the belief).

9

I don’t understand what you mean with

10

I want to replicate sample_prior_predictive manually. How would I do that with pm.draw?

11

You call pm.draw on all model variables


with pm.Model() as m:
  x = ...
  y = ...
  z = ...

pm.draw([x, y, z], draws=500)
# or
pm.draw(m.basic_RVs, draws=500)
1 Like