Matmul on deterministic variable TypeError

Questions v3
1

Hi there

I’m trying to run a non-centered hierarchical model, and running into a type error while trying to do matrix multiplication on a pm.Deterministic variable.

I have tried various different things like using np.matmul explicitly, using np.dot and taking the beta_unit * beta_sigma + beta_mu expression out of the pm.Deterministic class, but I run into errors with everything I try. (Happy to provide more details here).

This seems like relatively simple functionality; I wonder if anyone else has run into this, or if there’s a known way to achieve this?

Thanks in advance for any advice
David

import numpy as np
import pymc3 as pm

# random data to illustrate TypeError
X = np.random.randn(104, 216)
Y = np.random.rand(104)

# model
with pm.Model() as non_centered_model:
    
    # hyperpriors
    beta_mu = pm.Normal("beta_mu", mu=1, sigma=1)
    beta_sigma = pm.Exponential("beta_sigma", 1)
    
    # priors
    beta_unit = pm.Normal("beta_unit", mu=0, sigma=1, shape=X.shape[1])
    beta = pm.Deterministic("beta", beta_unit * beta_sigma + beta_mu)
    sigma = pm.Exponential("sigma", 1)
    
    # likelihood
    pm.Normal("lik", mu=X @ beta, sigma=sigma, observed=Y)
    
    idata_non_centered = pm.sample(return_inferencedata=True)

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Input In [150], in <cell line: 9>()
     18 sigma = pm.Exponential("sigma", 1)
     20 # likelihood
---> 21 pm.Normal("lik", mu=X @ beta, sigma=sigma, observed=Y)
     23 idata_non_centered = pm.sample(return_inferencedata=True)

TypeError: unsupported operand type(s) for @: 'numpy.ndarray' and 'DeterministicWrapper'
1 Like
2

When performing algebraic operations on parameters, you often have the functions in pm.math (many of which mirror their numpy counterparts). Also, see here.

3

Thanks for that. Sorry for not properly searching the docs.

1 Like