separating model definition from observations? about pymc2 HOT 6 CLOSED

I think something along these lines would be feasible. Perhaps the most sensible way would be to change Stochastic.set_value method to accept an observed flag.

umbrella.set_value(True, observed=True)

I'm not opposed to your solution either, however, where it is set via the model rather than the stochastic. No reason we can't have both, I suppose. All this would have to do is move stochastics to and from the Model.stochastics and Model.observed_stochastics dictionaries.

from pymc2.

Either way sounds great. I'd be happy to try to implement this. Should it go in pymc or pymc3? I'm not totally sure of the relationship between the two - is pymc3 meant to replace pymc very soon? Is the plan to have pymc3 be backwards compatible with pymc?

from pymc2.

You can try to implement it in either version, though pymc will be easiest. pymc3 will eventually replace pymc, but they are likely to coexist for some time. It is a breaking update; though the syntax is very similar, you cannot simply run a pymc model under pymc3.

from pymc2.

I would also be interested in this feature for similar reasons to Yarden-- I am training a model with a given set of observations, and then testing how well other observations fit that model. In particular, it would be really useful to have a function to evaluate the logp of a model given observations that are not extrinsic to the model itself. Perhaps these could even be optional arguments passed to the logp function

@stochastic(observed=True)
def my_distribution(**kwargs):
    #etc.

my_distribution.logp
#^^ would evaluate with values used at initialization
my_distribution.logp(other_vaues)
#^^ evaluates with values =  other_values instead of those used at initialization

from pymc2.

I would also love to have this feature! It would be nice to see how the posterior shifts as new data is observed.

from pymc2.

What I generally do is wrap my model in a function, and use the dataset as an argument. That way, each instance of the model is associated with the new dataset. e.g.

def generate_model(dataset):

    <model declaration>

    # Likelihood
    y = Normal('y', mu, tau, value=dataset['y'], observed=True)

    return locals()

model1 = MCMC(generate_model(dataset1))
model2 = MCMC(generate_model(dataset2))

from pymc2.