I have the following code,

```
with pm.Model(coords=coords) as baseModel:
# input data
xx_ = pm.MutableData("xx_", xx)
yy_ = pm.MutableData("yy_", yy_obs)
sm_ = pm.MutableData("sm_", sm_obs)
#priors
sigma_obs = 0.1
A_ = pm.Uniform('A_', 0, 1)
B_ = pm.Uniform('B_', 0, 1)
C_ = pm.Uniform('C_', 0, 1)
D_ = pm.Uniform('D_', 0, 1)
E_ = pm.Uniform('E_', 0, 1)
F_ = pm.Uniform('F_', 0, 1)
x1_ = A_ + B_*xx_
x2_ = C_ + D_*xx_
x3_ = E_ + F_*xx_
def f(x1_ = x1_, x2_ = x2_, x3_ = x3_, sm_ = sm_):
out = T11(x1_, x2_, x3_ ,sm_)
return out
function_pm = pm.Deterministic('s0f', f())
obs = pm.Normal('obs', mu=function_pm, sigma=sigma_obs, observed=yy_)
```

where T11 is an external function. As is, the model only samples the variables, `Sampling: [A_, B_, C_, D_, E_, F_, obs]`

. Why not x1, x2 and x3? I know I can recompute them, but I want to follow the scheme proposed here to use the posteriors of A… F once calibrated to do further inference, and the values of A…F are not stored in the posterior predictive.