Pymc3: ValueError: Unused step method arguments: {'target_accept'}

[ricardoV94]

Sampling fails when I specify target_accept

Code:

with pm.Model() as m:
    prior_mu = 30
    prior_sd = 15
    prior_size = -prior_mu**2 / (prior_mu - prior_sd ** 2)
    n_socks = pm.NegativeBinomial('n_socks', prior_mu, prior_size)

    prop_pairs = pm.Beta('prop_pairs', 15, 2)

    n_pairs = pm.Deterministic('n_pairs', pm.math.floor(pm.math.floor(n_socks / 2) * prop_pairs))
    n_odd = pm.Deterministic('n_odd', n_socks - (n_pairs * 2))

    post = pm.sample(500, target_accept=0.95)

Same error happens with:

    post = pm.sample(500, nuts_kwargs=dict(target_accept=0.95))

If I don't specify the target_accept, it runs fine:

Sequential sampling (2 chains in 1 job)
CompoundStep
>Metropolis: [n_socks]
>NUTS: [prop_pairs]
100%|██████████| 1000/1000 [00:00<00:00, 1376.44it/s]
100%|██████████| 1000/1000 [00:00<00:00, 1485.11it/s]
The gelman-rubin statistic is larger than 1.05 for some parameters. This indicates slight problems during sampling.
The estimated number of effective samples is smaller than 200 for some parameters.

Full traceback:

ValueError                                Traceback (most recent call last)

<ipython-input-21-f8cb4dc74d3d> in <module>()
     10     n_odd = pm.Deterministic('n_odd', n_socks - (n_pairs * 2))
     11 
---> 12     post = pm.sample(500, target_accept=.95)

2 frames

/usr/local/lib/python3.6/dist-packages/pymc3/sampling.py in sample(draws, step, init, n_init, start, trace, chain_idx, chains, cores, tune, progressbar, model, random_seed, discard_tuned_samples, compute_convergence_checks, **kwargs)
    404                 step = assign_step_methods(model, step, step_kwargs=kwargs)
    405         else:
--> 406             step = assign_step_methods(model, step, step_kwargs=kwargs)
    407 
    408         if isinstance(step, list):

/usr/local/lib/python3.6/dist-packages/pymc3/sampling.py in assign_step_methods(model, step, methods, step_kwargs)
    153             selected_steps[selected].append(var)
    154 
--> 155     return instantiate_steppers(model, steps, selected_steps, step_kwargs)
    156 
    157 

/usr/local/lib/python3.6/dist-packages/pymc3/sampling.py in instantiate_steppers(model, steps, selected_steps, step_kwargs)
     79     unused_args = set(step_kwargs).difference(used_keys)
     80     if unused_args:
---> 81         raise ValueError('Unused step method arguments: %s' % unused_args)
     82 
     83     if len(steps) == 1:

ValueError: Unused step method arguments: {'target_accept'}

Versions and main components

• PyMC3 Version: 3.7 and 3.8
• Theano Version:1.0.4
• Python Version: 3.6.9
• Platform: Google Colab

[AlexAndorra]

Hi Ricardo,
Yeah, it's because of the Compound Step, so you need to access NUTS step method directly, with a dict set to a lower case version of the step method's name.
Current docs are indeed unclear about that, but @jonsedar brought nice clarifications to the docstrings in a recent PR.

Basically, post = pm.sample(500, nuts={'target_accept':0.95}) should do the trick.
Tell me if this works, and if yes I'll close 🖖

[ricardoV94]

Yes, that solves it!

[jonsedar]

Indeed - see the diff here :) https://github.com/pymc-devs/pymc3/pull/3908/files#diff-7eb6c4a83cfe45b9fc0eac76b57e2175