sample_prior_predictive and sample_posterior_predictive issues with use_auto_batching=False

[dirmeier]

Hi,

when sampling from the prior predictive with use_auto_batching=False setting sample_shape doesn't work and always a single draw is returned:

dat = tfd.Normal(loc=1.0, scale=1.0).sample(10)

@pm.model
def model():
    m = yield pm.Normal("mean", loc=0.0, scale=1.0)
    lik = yield pm.Normal("lik", loc=m, scale=1.0, observed=dat)
    return lik

mod = model()
ppc = pm.sample_prior_predictive(mod, sample_shape=10, use_auto_batching=False)
ppc = ppc.prior_predictive
print(ppc["model"])

<xarray.DataArray 'model' (chain: 1, draw: 1)>
array([[-0.6240528]], dtype=float32)
Coordinates:
  * chain    (chain) int64 0
  * draw     (draw) int64 0

Similar behaviour is true when sampling from the posterior predictive. Am I missing something here? Colab that shows this is here.

Thanks!

Cheers,
Simon

[lucianopaz]

@dirmeier, the problem is that you didn't specify that m is conditionally_independent=True. This isn't properly documented because I never had the time to write an explanatory notebook on how to manually vectorize a model. The closest thing there is at the moment is this example from the forward sampling functions

[dirmeier]

Ah, right! Thanks.