Errors like this might be due to the random character of MCMC sampling, which might occasionally get stuck (in which case it would not be system-specific, but coincidence; see suggestion below). However, I guess you tried more than once.
In your definition of p, you use the shape parameter:
Why do you set shape=2000? This creates an “array” (sorry if the term is inappropriate) of 2000 “pm.Beta” objects.
instead, I guess you want one ‘p’, right?
Best,
Falk
To compare across systems, you can try fixing the random state in numpy and pymc3/theano, so that the sampling trace should be identical on both machines:
- choose a number (
seed) - set
np.random.seed(seed)just after the import ofnumpy - sample with:
trace = pm.sample(random_seed = seed)