I struggle to debug this seemingly simple model: with pm.Model() as model: ratio = pm.Dirichlet('ratios', a=alpha, testval=start) obs = pm.Multinomial('observations', n=total, p=ratio, observed=local_counts) trace = pm.sample(draws=1000, cores=8) It fails with ValueError: Bad initial …

If all(np.isfinite(q0)) is False, it would means some starting value is non-finite. Sometimes it might not be visible in the test point, as in this case the free parameters are actually the transformed version of the Dirichlet. Which means internally, the test_point is transformed (on to the real …

Try trace = pm.sample(draws=1000, cores=8, init='adapt_diag') The default jitter might put the starting values to a invalid point.

It still fails similarly: Auto-assigning NUTS sampler... INFO:pymc3:Auto-assigning NUTS sampler... Initializing NUTS using adapt_diag... INFO:pymc3:Initializing NUTS using adapt_diag... Multiprocess sampling (8 chains in 8 jobs) INFO:pymc3:Multiprocess sampling (8 chains in 8 jobs) NUTS: [ratios] I…

I ran the steps and got {'ratios_stickbreaking__': array([ 0.98842164, -3.49124823, -0.05789614, ..., -7.64635372, -8.74488638, 2.23058342])} [ 0.33077722 -0.60774967 -0.96290679 ... -0.89012621 -0.32552028 -0.10242498] inf -inf [-1.68691344 0.96953555 0.05625001 ... 0.99936306 0.99976…

You are right. I investigated the test value for the Dirichlet distribution and saw that all(start[:-1][~np.isfinite(q0)] == 0) is True. It does not really make sense to have any p_i to be exactly zero. So I chose a nicer start value: start = local_counts.sum(axis=0) + 1 start = start / start.sum()…

Large Feature Number Dirichlet

Questions

junpenglao December 5, 2018, 2:05pm 6

Sounds like there is something wrong with the gradient - try following the steps in How to track a 'nan energy'?

Topic		Replies	Views
SamplingError: Bad initial energy Questions	1	2826	September 10, 2020
ValueError: Bad initial energy: inf. The model might be misspecified error replicating Multilevel Regression and Poststratification with PyMC3 notebook Questions	2	1640	May 11, 2018
Dirichlet Distributions and Batching in pymc3 3.9+ Questions	2	493	February 1, 2021
Fixing ValueError: Bad initial energy: inf. The model might be misspecified Questions	2	502	October 30, 2018
Bad initial energy after checking test point Questions	4	652	December 6, 2020

Large Feature Number Dirichlet

Related topics