CODEWITHSUNDEEP
pythonlinear-regressionpymcpymc3
jonnydedwards
jonnydedwards

Reputation: 180

Pymc3: very slow and stalling

is there any reason why the NUTS sampler might be slow or stall? I'm using http://twiecki.github.io/blog/2014/03/17/bayesian-glms-3/ as a basis for some hierachical linear regression work. I've tried starting with find_MAP() and it's still stalls after 100 of the 2000 iterations

my code is 

with pm.Model() as hierarchical_model:
   # Hyperpriors for group nodes
   mu_a = pm.Normal('mu_alpha', mu=0., sd=100**2)
   sigma_a = pm.Uniform('sigma_alpha', lower=0, upper=100)
   mu_b = pm.Normal('mu_beta', mu=0., sd=100**2)
   sigma_b = pm.Uniform('sigma_beta', lower=0, upper=100)

   a = pm.Normal('alpha', mu=mu_a, sd=sigma_a, shape=n_dis)
   b = pm.Normal('beta', mu=mu_b, sd=sigma_b, shape=n_dis)

   # Model error
   eps = pm.Uniform('eps', lower=0, upper=100)

   actual_est = a[disRefV] + b[disRefV] * data.baseline.values

   actual_like = pm.Normal('actual_like', mu=actual_est, sd=eps, observed=data.prepanel)

with hierarchical_model:
    start = pm.find_MAP()
    step = pm.NUTS()
    hierarchical_trace = pm.sample(2000, step, progressbar=True)

many thanks in advance

Upvotes: 3

Views: 2048

Answers (1)

twiecki
twiecki

Reputation: 1316

NUTS does get stuck sometimes. Did you z-score / standardize your data? That often helped in my experiments. If you are using a hierarchical model you'll probably want to z-score your data according to the group mean, not the individual means.

You can then rescale your posteriors so that they are in the original space (formula 17.1 in the Kruscke book "Doing Baysian Data Analysis").

Upvotes: 3

Related Questions