Hi, I’m trying to use Pymc3 to find the appropriate number of componements used to predict a time serie of concentrations called aCH and aCOH. These are two distincts PLS models. Each model can use between 1 and 20 number of components. The predictions are stored in an ndarray with the time series …

Yes using Metropolis could produce invalid proposal as described in the other post you mentioned, but using Categorical distribution seems to work: n = aCH_.eval().shape[1] with pm.Model() as basic_model: # Priors for unknown model parameters b1 = pm.Uniform('b1', lower=0.3, upper=0.5, test…

I have run this model with 8 millions iterations with 1 millions tuning steps and still have a Gelman Rubin above 1.4 and a lot of divergences. Do you know have any idea of what causes that problem? Increasing target_accept doesn’t help either. [Capture]

More iterations doesn’t help when there are problems in your model. In this case, a few suggestion: Increase target_accept leads to smaller step size, but it only helps when you are seeing <10 divergences. When there are lots of divergent, it means something is wrong with the model looking at the …

Thanks for the advices! I’m going to try to mix Categorical GibbsSampler with Metropolis instead of Nuts as these sampling methods are closer. Concerning the b1 coefficient, it really should be a scaler and not a vector. It is a physical property that should be the same for each sample.

If I use Metropolis, then the variables updated by Categorical Gibbs Sampler are not updated at all. This is very strange. Do you have any idea why?

I tried. Doesn’t move from the starting values

I had another look at the model, with the following code: with pm.Model() as basic_model: # Priors for unknown model parameters b1 = pm.Uniform('b1', lower=0.3, upper=0.5, testval=0.45) ncomp_aCH = pm.Categorical('ncomp_aCH', p=np.ones(n)/n) ncomp_aCOH = pm.Categorical('ncomp_aCOH',…

I think there is a way to marginalized instead of indexing. The thought behind it is that when you computed the posterior expectation of ncomp_aCH or ncomp_aCOH, it’s a weight vector with n elements that summed to 1. So instead of restricting the vector to take value only 0 or 1 (ie an index vector)…

This sounds like a very intersting option. I will give it a try, will come back to you. Thank you!

Indexing using a free random variable

Questions

junpenglao October 10, 2019, 8:08am 25

Looks like you are trying to do latent Dirichlet allocation. You should try to reparameterized your model into a marginalized version: Frequently Asked Questions

Topic		Replies	Views
Marginalizing out a categorical variable Questions	18	2321	April 1, 2021
Array index generated by Categorical Distribution Questions theano	2	1240	June 17, 2018
Problem with DiscreteUniform giving indices out of bounds Questions	6	570	March 4, 2021
Inferring Dirichlet concentration parameter Questions	11	1321	May 13, 2020
Marginal likelihood for distributions with discrete variables v3	12	1567	November 14, 2022

Indexing using a free random variable

Related topics