Sampling error: Random variables detected in the logp graph

Questions v5
1

Dear all,

I’m trying to implement the ALCOVE model(attention learning model for category learning) using PyTensor operations. The model code is shown below.

with pm.Model() as model:
    ##########Priors##########
    c = pm.Uniform("c", lower=0, upper=10)
    phi = pm.Uniform("phi", lower=0, upper=10)
    alr = pm.Beta("alr", alpha=1, beta=1)
    wlr = pm.Beta("wlr", alpha=1, beta=1)
    
    #########alcove simulation probs mean by Block#####
    #p_block = pm.Deterministic("p_block", block_probs_from_theta(c, phi, alr, wlr))
   
    #########Binomial Likelihood#########
    y = pm.Binomial("y", n=n_block, p=p_block, observed = y_block) 
    idata = pm.sample()

print(az.summary(idata, var_names=["c", "phi", "alr", "wlr", "p_block"]))
az.plot_trace(idata, var_names=["c", "phi", "alr", "wlr"])

However, I get the following error:

ValueError                                Traceback (most recent call last)
Cell In[2], line 80
     75     #print(p_block.eval().shape)
     76 
     77     #########Binomial Likelihood#########
     78     y = pm.Binomial("y", n=n_block, p=p_block, observed = y_block) 
---> 80     idata = pm.sample()
     82 print(az.summary(idata, var_names=["c", "phi", "alr", "wlr", "p_block"]))
     83 az.plot_trace(idata, var_names=["c", "phi", "alr", "wlr"])

File ~\anaconda3\envs\pymc_env\lib\site-packages\pymc\sampling\mcmc.py:679, in sample(draws, tune, chains, cores, random_seed, progressbar, step, var_names, nuts_sampler, initvals, init, jitter_max_retries, n_init, trace, discard_tuned_samples, compute_convergence_checks, keep_warning_stat, return_inferencedata, idata_kwargs, nuts_sampler_kwargs, callback, mp_ctx, model, **kwargs)
    676         auto_nuts_init = False
    678 initial_points = None
--> 679 step = assign_step_methods(model, step, methods=pm.STEP_METHODS, step_kwargs=kwargs)
    681 if nuts_sampler != "pymc":
    682     if not isinstance(step, NUTS):

File ~\anaconda3\envs\pymc_env\lib\site-packages\pymc\sampling\mcmc.py:210, in assign_step_methods(model, step, methods, step_kwargs)
    208 methods_list: list[type[BlockedStep]] = list(methods or pm.STEP_METHODS)
    209 selected_steps: dict[type[BlockedStep], list] = {}
--> 210 model_logp = model.logp()
    212 for var in model.value_vars:
    213     if var not in assigned_vars:
    214         # determine if a gradient can be computed

File ~\anaconda3\envs\pymc_env\lib\site-packages\pymc\model\core.py:717, in Model.logp(self, vars, jacobian, sum)
    715 rv_logps: list[TensorVariable] = []
    716 if rvs:
--> 717     rv_logps = transformed_conditional_logp(
    718         rvs=rvs,
    719         rvs_to_values=self.rvs_to_values,
    720         rvs_to_transforms=self.rvs_to_transforms,
    721         jacobian=jacobian,
    722     )
    723     assert isinstance(rv_logps, list)
    725 # Replace random variables by their value variables in potential terms

File ~\anaconda3\envs\pymc_env\lib\site-packages\pymc\logprob\basic.py:631, in transformed_conditional_logp(rvs, rvs_to_values, rvs_to_transforms, jacobian, **kwargs)
    629 rvs_in_logp_expressions = _find_unallowed_rvs_in_graph(logp_terms_list)
    630 if rvs_in_logp_expressions:
--> 631     raise ValueError(RVS_IN_JOINT_LOGP_GRAPH_MSG % rvs_in_logp_expressions)
    633 return logp_terms_list

ValueError: Random variables detected in the logp graph: {wlr, phi, c, alr}.
This can happen when DensityDist logp or Interval transform functions reference nonlocal variables,
or when not all rvs have a corresponding value variable.

All model computations are defined inside a function called block_probs_from_theta.
If I manually assign constant values to the parameters (e.g., c=1.0, phi=1.0, alr=0.5, wlr=0.5), the function runs fine and produces the expected output.
But when I use these parameters as PyMC random variables inside the model, the sampling fails.

What does this error mean, and how can I properly connect my PyTensor function with the PyMC model so that sampling works?

2

Doesn’t block_probs_from_theta use random variables?

3

Ah are you commenting p_block, that means p_block is probably referencing a variable from another PyMC model (like the one you ran before commenting out, if you are in an interactive notebook).

Way to replicate this error:

import pymc as pm

with pm.Model() as m:
  x = pm.Normal("x")

with pm.Model() as m:
  y = pm.Normal("y", mu=x)
  pm.sample()

Note the second model m is using x, which is not defined in the model itself. That will complain. We’ve changed the error in recent versions of PyMC, and it would output the following:

ValueError: Random variables detected in the logp graph: {x}.
This can happen when mixing variables from different models, or when CustomDist logp or Interval transform functions reference nonlocal variables.
4

So the error you’re reporting here is probably not the error you had originally and are interested in debugging, just an accidental error from commenting out that line.

5

Oh, In my actual code, the p_block line was not commented out when running the model. It was a typo when posting :sweat_smile: block_probs_from_theta only use random variables defined in the model.

6

If you’re seeing the same error, it still sounds like referencing variables not from the model (or creating new random variables with .dist())