Number of samples is high when init='nuts'

[rosgori]

Description of your problem

Please provide a minimal, self-contained, and reproducible example.

When you use the option init='nuts' in pm.sample(), the number of samples is really high

import pymc3 as pm
import numpy as np

N = 1000
x = np.random.normal(loc=0, scale=2, size=N)

with pm.Model() as model:
    
    mu = pm.Normal('mu', 0, 5)
    sigma = pm.HalfNormal('sigma', sd=5)
    
    obs = pm.Normal('obs', mu=mu, sigma=sigma, observed=x)
    
with model:
    trace = pm.sample(draws=100, tune=100, init='nuts')
    
print(pm.summary(trace))

Please provide the full traceback.

Sampling 4 chains, 0 divergences:   1%|▏         | 15137/1200000 [00:05<07:25, 2661.65draws/s]

As you see, the number 1200000 always appear, it doesn't matter if you change draws or tune in pm.sample.

Should that number change when you change draws and tune?

Versions and main components

• PyMC3 Version: 3.8
• Theano Version: 1.0.4
• Python Version: 3.7.6
• Operating system: Ubuntu 16.04.6 LTS
• How did you install PyMC3: (conda/pip) conda

[twiecki]

Thanks, that init method should just be dropped. Where did you find it referenced?

[AlexAndorra]

It seems to be in pm.sample doc:

nuts: Run NUTS and estimate posterior mean and mass matrix from the trace.

Then, in the source code, when the init method is "nuts", it samples with n_init as the number of draws:

 elif init == "nuts":
      init_trace = pm.sample(
          draws=n_init, step=pm.NUTS(), tune=n_init // 2, random_seed=random_seed
      )

Since pm.sample defaults to n_init=200_000, this yields 200_000 + 100_000 samples per chain, which is, I think, where the 1_200_000 come from (since you're sampling 4 chains in the example).

[twiecki]

Yeah this is a left-over, we have better ones now so should just remove it.

[AlexAndorra]

Ok, seems straightforward -- I'll do it

[Ahanmr]

@twiecki @AlexAndorra This issue is fixed right? Is there still anything pending on this one? Looks like init='nuts' is mostly dropped already.