# Can't sample a Beta(70,70) ! (but a beta(60,60) works ?!)

While trying some experiment about beta distributions I found thid weird behaviour…

This code works :

``````with pm.Model() as model:
r0 = pm.Beta('r0',alpha=1+60,beta=1+60)

trace = pm.sample(chains=1)
_ = pm.traceplot(trace)
``````

,
But this one, just changing the alpha & beta params from 60 to 70 :

``````with pm.Model() as model:
r0 = pm.Beta('r0',alpha=1+70,beta=1+70)

trace = pm.sample(chains=1)
_ = pm.traceplot(trace)
``````

fails with a : “ValueError: Bad initial energy: nan. The model might be misspecified.”

I don’t understand why !? And this is even more strange :
`r0 = pm.Beta('r0',alpha=1+999,beta=1+999)`
works !!

Why certain specific parameter values range seem to have problem ?

Any help appreciated.

Hm, I can’t seem to reproduce this, that’s really interesting.

Perhaps “bad initial energy” means that the sampler is trying to start in a region of zero probability density. Since this Beta has such tight tails, could it be that an initialization outside roughly (0.3, 0.7) is causing that problem?

Interesting hypothesis, but then :

`r0 = pm.Beta('r0',alpha=1+999,beta=1+999)`

Should be even harder to sample !! (but it works perfectly…)

Yeah, I can’t explain that one Perhaps there’s randomness in the initialization? I may just be making that up, though, I could be completely wrong here.

Can you rerun the Beta(70, 70) and see if it still fails?

1 Like

I’ve restarted my computer, and I still have this weird behavior !

[…]

And if I change the beta distribution parameters to : (60,60)… it works fine :

I have no clue !! Except that it seems not to happen on a another computer (thank @tushar_chandra), which is a nice thing, but I still find this worrying for a futur use of pymc3 in production…)

I have not been able to reproduce. Have you tried using different random seeds when calling `pm.sample`? Or set a different `start` and `init` combination?

No seed where specified , so I guess the starting point is different for each run.
(I run both version several time with the same weird behavior…)

I have the final answer for that bug !!

I had the latest numpy installed instead of numpy-1.14.3 (recommanded for pymc3).

uninstall numpy then re-install numpy-1.14.3, and this bug desappears…

1 Like