Re-working `as_model` about pymc-experimental HOT 10 CLOSED

Definitely the second, I don't think a user should have to remember to call the helper. In the previous we also deal with coords transparently.

from pymc-experimental.

Hmm, I think I'd rather keep the one model. Unless my current workflow is bad practice or I've missed the point:

m = model(data=data, coords=coords)
idata = pm.sample(
    model=m,
    return_inferencedata=True,
)

idata.extend(pm.sample_posterior_predictive(idata, model=m, predictions=False))

pm.set_data(new_data=data_predict, model=m)
idata.extend(pm.sample_posterior_predictive(idata, model=m, predictions=True))

tbf, I guess replacing the set_data line isn't a killer. Might be easier to keep track of one m though. Coming from R (sorry), I prefer one m: predict(m, newdata = new_data)

With either of the above workflows your code can look like this:

m = model(data=data, coords=coords)
idata = pm.sample(model=m)

pm.sample_posterior_predictive(idata, model=m, predictions=False, extend_inferencedata=True)
m_new = model(data=data_predict, coords=new_coords)
pm.sample_posterior_predictive(idata, model=m_new, predictions=True, extend_inferencedata=True)

from pymc-experimental.

I prefer the second. As long as it works for coords_mutable (and whatever other args we need) too.

from pymc-experimental.

You may have no need for coords_mutable with as_model since you can just build a different model with fixed coords for posterior predictive, as shown above with m_train and m_test

from pymc-experimental.

Hmm, I think I'd rather keep the one model. Unless my current workflow is bad practice or I've missed the point:

m = model(data=data, coords=coords)
idata = pm.sample(
    model=m,
    return_inferencedata=True,
)

idata.extend(pm.sample_posterior_predictive(idata, model=m, predictions=False))

pm.set_data(new_data=data_predict, model=m)
idata.extend(pm.sample_posterior_predictive(idata, model=m, predictions=True))

tbf, I guess replacing the set_data line isn't a killer. Might be easier to keep track of one m though. Coming from R (sorry), I prefer one m: predict(m, newdata = new_data)

from pymc-experimental.

I'm starting to appreciate @ricardoV94 's approach of simply having a prediction model that uses the trace of the fitted model, and not have to worry about changing coords.

I'm confused by the example above -- m_train and m_test appear to be two entirely separate model instances, with no shared parameters. How is m_test getting the trained parameters?

from pymc-experimental.

How is m_test getting the trained parameters?

That would happen when calling pm.sample_posterior_predictive with the idata from sampling

from pymc-experimental.

So the missing steps are:

trace = pm.sample(model=m_train)
pm.sample_posterior_predictive(trace, model=m_test, extend_inferencedata=True)

The nice thing here is that it avoids having to set up placeholder vars when you build a separate prediction model.

Yeah, I like the second one.

from pymc-experimental.

Still in favor of the second.

from pymc-experimental.

Thanks everyone for your feedback!

from pymc-experimental.