Is there a way to represent a model as a DAG in Pymc3?

[bpajusco]

I liked the the graph generation functionality provided by PyDot in Pymc 2 and assumed it would have ported to Pymc3 so I tried to import 'graph' from pymc and got a cannot import module 'graph' error. The error message is self explanatory but makes me wonder whether we have that possibility with Pymc3. Given that is based on Theano I'm inclined to believe that is possible to generate a graph of the model but I have no idea how to. Is there a way to learn about this?

Thanks,
Bruno

[fonnesbeck]

There is no DAG graphing currently in PyMC 3, but it is something we need. I know theano has a FunctionGraph object, and others have devised ways of turning these into DAGs. I would be inclined to incorporate igraph to do the graphical side of this.

[datnamer]

would this project be helpful? http://daft-pgm.org/

[majidaldo]

wouldn't this be just following the parent/child relationships of the nodes?

[twiecki]

@majidaldo Yes, but we don't have that relationship encoding anywhere right now.

[fonnesbeck]

You can certaintly use Theano's pydotprint to get a DAG of the Theano graph, but you might not like the look of it:

with pymc3.Model() as basic_model:
    # Priors for unknown model parameters
    sigma_beta = pymc3.HalfNormal('sigma_beta', 0.5)
    alpha = pymc3.Normal('alpha', mu=0, sd=1)
    sigma = pymc3.HalfNormal('sigma', 1.5)
    beta = pymc3.Normal('beta', mu=0, sd=sigma_beta, shape=2)

    mu = alpha + beta[0] * X1 + beta[1] * X2

    # Likelihood (sampling distribution) of observations
    Y_obs = pymc3.Normal('Y_obs', mu=mu, sd=sigma, observed=Y) 

from theano.printing import pydotprint

pydotprint(basic_model.logpt)

[twiecki]

#1683