Description
Prior predictive sampling draws values from the untransformed variables and then attempts to recreate the transformed values via transformation.forward_val(untransformed_values)
However, this logic is not sufficient to recover the correct values when there are stochastic bounds (e.g., pm.Uniform('x', lower=0, upper=y), because we also need to know what the values of the stochastic bounds were in each sample (i.e., the y).
People don't usually care about transformed variables, but this is critical for example in sample_smc() as it creates it's particles from an initial prior_predictive call.
Hopefully this will be gone in V4.0, but I thought best to document it just in case (and in the meanwhile).
Minimal reproducible example:
np.random.seed(1)
with pm.Model() as m:
y = pm.Uniform('y', 0, 1)
x = pm.Uniform('x', 0, y)
prior = pm.sample_prior_predictive()
print(np.mean(np.isnan(prior['x_interval__'])))/home/ricardo/Documents/Projects/pymc3/pymc3/distributions/transforms.py:294: RuntimeWarning: invalid value encountered in log
return floatX(np.log(x - a) - np.log(b - x))
0.314
The 31.4% NANs correspond to cases where a and b were either above or below x due to random sampling in the block below. The NANs are the most obvious symptom, but all values are in fact incorrect:
Which is called at the end of prior_predictive_sampling, without providing a point or any other contextual info: