Resampling observations from a model instance

[phipsgabler]

Capturing this from Slack: appearently, there's currently no easy way to sample from the generative process via an instantiated model.

The evaluator closure internally just performs a little runtime check on the arguments to decide whether to something is not in the arguments, or missing, so you could just create a new model with the same f and defaults, but modified args.

@torfjelde suggested to change isassumption to a function that takes contexts (which would in any case be much more elegant than the current version), and provide a new context which also resamples the observations (surely that could use the same missing interface as generative sampling, so that you can also partially regenerate observations).

[phipsgabler]

@torfjelde we can do this with decondition now, right? Does that also work for observations passed as arguments?

[devmotion]

Usually I pointed out in such discussions that one can specify the missings type parameter if one does not want to pass fake/missing arguments:

DynamicPPL.jl/src/model.jl

Line 44 in ad545be

Model{missings}(name::Symbol, f, args::NamedTuple, defaults::NamedTuple)

[torfjelde]

Yeah. My approach would now be:

for batch in batches(data)
    cmodel = condition(model, data=batch)
    # Do whatever you want.
    ...
end

[torfjelde]

So something like Stochastic HMC would just be:

using Turing, Random, DynamicPPL

@model function demo(n=1, ::Type{TV} = Vector{Float64}) where {TV}
    m ~ Normal()
    x = TV(undef, n)
    for i in eachindex(x)
        x[i] ~ Normal(m, 1)
    end
end

data = randn(100)
batch_size = 10

model = demo(batch_size)

spl = DynamicPPL.Sampler(HMC(1e-3, 10))

# Initialize.
transition, state = let batch = data[1:batch_size]
    AbstractMCMC.step(rng, condition(model, x=batch), spl)
end

ts = AbstractMCMC.samples(transition, model, spl, 1)

# Iterate over all the batches.
for batch in Iterators.partition(shuffle(rng, data), batch_size)
    cmodel = condition(model, x=batch)
    transition, state = AbstractMCMC.step(rng, cmodel, spl, state)
    AbstractMCMC.save!!(ts, transition, length(ts) + 1, model, spl)
end

[torfjelde]

Does that also work for observations passed as arguments?

But no, it won't work for these cases. For this you need to work with the missings as @devmotion said, but I really don't like that approach tbh 😕

[phipsgabler]

Alright, so can we close this issue and refer to condition/decondition if the question comes up in the future?

[torfjelde]

On the DynamicPPL-side I think this can be closed. I do think we should maybe have some functionality for making this easier on the user, but that should probably go in Turing.jl or something.

[torfjelde]

I realized I completely misunderstood the question, haha. But yeah, you can do it using decondition now if you use that interface.

[phipsgabler]

Oh... and I was wondering about your answers and thought those examples were some special use case that I just didn't get 😅