Weighted Inference with Multivariate Normal Likelihood

1

Hi,

I am aiming to perform MCMC sampling with a weighted multivariate normal likelihood, similar to this topic, but I receive an error when trying to recreate the weighted likelihood function in the sampling. Does anyone know how I might resolve this? My code and the error are below:

model = pm.Model()

with model:

    #Prior
    mu = pm.MvNormal('mu', mu=[0,0], cov=5*np.identity(2))

    #covariance
    cov = np.identity(2)

    #weighted likelihood
    weights = np.ones(len(data))
    weighted_likelihood = pm.Potential('weighted_likelihood', weights*pm.MvNormal.dist(mu=mu,cov=cov).logp(data))

    #MCMC Posterior sampling
    idata = pm.sample(mp_ctx='spawn')

And the error doesn’t recognise format of the the MvNormal.dist() inputed into the pm.Potential function as found in the original topic:

AttributeError: 'TensorVariable' object has no attribute 'logp'

If anyone has any tips that might help it would be appreciated!

2

.logp is the syntax old PyMC used to have. It’s been replaced by pm.logp(pm.MvNormal.dist(...), data)

1 Like