ERROR [pymc3] The rhat statistic is larger than 1.4 for some parameters. The sampler did not converge.

[wmf1991yeah]

A problem exists: "ERROR [pymc3] The rhat statistic is larger than 1.4 for some parameters. The sampler did not converge", when I run the PyMC3 sampleing using SEIR model. The problem as follow:

How to address the problem?

PS: The model have a larger number of parameters because more change points are set (8 change points). NUTS: [sigma_obs, delay_log, incub, I_begin, new_E_begin, mu, transient_len_7log, transient_len_6log, transient_len_5log, transient_len_4log, transient_len_3log, transient_len_2log, transient_len_1log, transient_day_7, transient_day_6, transient_day_5, transient_day_4, transient_day_3, transient_day_2, transient_day_1, lambda_7log, lambda_6log, lambda_5log, lambda_4log, lambda_3log, lambda_2log, lambda_1log, lambda_0log]

[semohr]

The easiest way to solve this is probably to increase the burn in steps and the sample steps.

For example the daily running scripts on the old repository have a burn in of 1000 steps and also 1000 samples.

I think you can figure that out ;)

This error can also occure if your priors are not set correctly, but one can only (as far as I know) test this by trial and error.

[wmf1991yeah]

Thank you very much! I try to add the number of steps and samples. Our priors is default distribution in the paper. In addtion, Is the PyMC3 support GPU? Can pymc3 run on the GPU?

[semohr]

I doubt the rhat error is caused by your priors, test it with more samples/burnin first.

Pymc3 uses theano as backend. Theano supports gpu calculation but if you do not have special hardware for that I doubt I will give you any perfomance increases.

Havent tested that though, feel free to update us on that if you get it running.

[wmf1991yeah]

Thank you very much! I try it using more samples (tunes=500, draws=1000, chains=2), then I achieve better result. I don't learn to how adding GPU support in pyMC3 although I have GPU hardware (GTX 1080Ti).

[jdehning]

The Hamiltonian Monte Carlo algorithm is inherently sequential. So you won't have an advantage by using a GPU

[wmf1991yeah]

OK, I see.