Description
Describe the issue:
When using coords in conjunction with Data some unexpected behavior can happen if one of the data containers is given the same name as one of the existing coords dimensions. The symptom I was diagnosing was pretty difficult to identify: when sampling posterior predictive checks, some variables in the middle of the model that used one of the coord dimensions in their dims were getting sampled for seemingly opaque reasons compared to identically constructed models in a separate notebook. Eventually realized that it was because of the conflicting namespacing, which I think led to the model thinking there was a shared dependency and unknown data leading to using the priors in the ppc. All would have been avoided if there was a simple warning when your Data container conflicts with a coord dim.
In the example below, changing the data container name to anything other than “group” is sufficient to return to normal behavior.
Reproduceable code example:
import numpy as np
import pymc as pm
import arviz as az
# Generate synthetic data
n_per_group = 100
groups = np.array([0, 1, 2])
group_means = np.array([0, 10, 20])
# Generate data for each group
data = np.concatenate([np.random.normal(loc=mean, scale=1, size=n_per_group) for mean in group_means])
group_codes = np.concatenate([[group] * n_per_group for group in groups])
# Define coordinates
coords = {"group": groups,}
with pm.Model(coords=coords) as model:
# Data containers
group = pm.Data("group", group_codes, dims="obs")
y = pm.Data("y", data, dims="obs")
# Priors
mu = pm.Normal("mu", mu=0, sigma=10, dims="group") # One mu for each group
sigma = pm.HalfNormal("sigma", sigma=1)
# Likelihood
y_obs = pm.Normal("y_obs", mu=mu[group], sigma=sigma, observed=y, dims="obs")
# Sampling
trace = pm.sample(1000,tune=1000,chains=1,return_inferencedata=True)
ppc = pm.sample_posterior_predictive(trace)Error message:
No response
PyMC version information:
5.18.2
Context for the issue:
It’s a footgun. It’s natural to want to use the same name for the dimension in coords and for the data that stores the assigned values, especially when you are dynamically constructing your model (eg from data frames which may have an unknown number of columns).
a simple warning (or hard error) which is thrown when this conflict occurs would be much appreciated.