Get Turing's model predictions after optimization

Specific Domains Probabilistic programming
1

Hi, since Turing 0.35 generated_quantities() is deprecated. So how do I now get my model’s predictions? Before 0.35 I did something like this

function predict(mdl, mles)
    ## mdl is an instantiated Turing model
    params = string.(names(mles.values)...)
    preds = generated_quantities(mdl, mles.values, params)
    return preds
end

I could write a function like expectation(params, data) that I use inside the model to get the expectation and then after fit to get the predictions, but that is rather cumbersome I feel. Is there an easier solution?

1 Like
2

Checkout the returned function described here.

3

Thanks, but I need to sample for this to work, if I understand correctly. However, I don’t sample, I do mode estimation as described here

4

I see the problem now. There is a method that accepts a NamedTuple as an input for the parameters. One hack would be to convert your mode estimation result into a NamedTuple. As a long term solution, maybe a method could be created to work specifically with mode estimators.

1 Like
5

Your predict function doesn’t work if you replace generated_quantities with returned? It looks like it’s the same. returned uses fix on the model and then calls it so this would be another way to write it.

6

Thanks to both of you, it works.

function predict(mdl, mles)
    ## mdl is an instantiated Turing model
    vs = Tuple(mles.values)
    ks = Tuple(names(mles.values)[1])
    preds = returned(mdl, NamedTuple{ks}(vs))
    return preds
end
1 Like
7

The API of the mode estimation bit in Turing is a bit undeveloped compared to the MCMC side I think. Will put this on the list :slight_smile: create `returned()` method for results from mode estimation · Issue #2607 · TuringLang/Turing.jl · GitHub