SamplingError: Initial evaluation with High value of N trials in binomial distribution

Questions
1

Im trying to evaluate A/B test results following this What is A/B testing? — PyMC3 3.11.4 documentation

I have been strugling with the simulation of Value conversions. Got a lot of SamplingError. I thought that was from the gamma unadapted choice of priors. But i have found out that even for simple Bernouilli conversions evaluation it failed. Whatever the prior’s choices for the Beta passed in the Binomial i had the same error. I also tried to play with the number of chains and cores but still got same error.

I have finaly found out that with toy data with small trials it works perfectly fine. As soon as i got high N passed it fails. I even used the small example from the sample method in the pymc doc (pymc.sampling — PyMC dev documentation). I still observe the same behavior. I tried pymc3 3.11.4 and 3.11.2 with same results.

Any ideas whats going wrong ?

Thanks all !


# example for pymc3 documentation

import pymc3 as pm

# Using a ratio high makes the sampling fail

ratio = 1000

# With small ratio works fine

# ratio = 1

n = 100*ratio

h = 61*ratio

alpha = 2

beta = 2

with pm.Model() as model: # context management

    p = pm.Beta("p", alpha=alpha, beta=beta)

    y = pm.Binomial("y", n=n, p=p, observed=h)

    trace = pm.sample()

---------------------------------------------------------------------------

SamplingError                             Traceback (most recent call last)

/tmp/ipykernel_30286/3176206272.py in <module>

      9     p = pm.Beta("p", alpha=alpha, beta=beta)

     10     y = pm.Binomial("y", n=n, p=p, observed=h)

---> 11     trace = pm.sample()

/data/anaconda/envs/py37_new/lib/python3.7/site-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, callback, jitter_max_retries, return_inferencedata, idata_kwargs, mp_ctx, pickle_backend, **kwargs)

    426     start = deepcopy(start)

    427     if start is None:

--> 428         check_start_vals(model.test_point, model)

    429     else:

    430         if isinstance(start, dict):

/data/anaconda/envs/py37_new/lib/python3.7/site-packages/pymc3/util.py in check_start_vals(start, model)

    238                 "Initial evaluation of model at starting point failed!\n"

    239                 "Starting values:\n{}\n\n"

--> 240                 "Initial evaluation results:\n{}".format(elem, str(initial_eval))

    241             )

    242

SamplingError: Initial evaluation of model at starting point failed!

Starting values:

{'p_logodds__': array(0., dtype=float32)}

Initial evaluation results:

p_logodds__   -0.98

y              -inf

Name: Log-probability of test_point, dtype: float64
2

Your code works fine on my side (pymc3 version 3.11.1)… I even increased ratio a couple of orders of magnitude and it still worked.

No idea of what could be going on.