[feature request] one-arg tempered_evaluate

[miguelbiron]

Hey @sunxd3 -- we realized that working with flattened_values as state for our samplers is going to result in loss of precision for models with mixed states. For example, when we have ints and floats, flattened_values is concretely typed Vector{Float64} so the ints are cast as floats. And that ends up causing issues when we want to dispatch for coordinate-wise sampling. Therefore, it would be nice to just work with the evaluation_env inside the model as the proper state. But then we don't have a way of computing tempered_evaluate!!(model).

Going further, it seems that (in the long run) it would make your life easier to have a unique "barebones_tempered_evaluate!!" such that all the special cases -- with/without tempering, with/without rng, and with/without flattened values -- are handled via light pre-processing of inputs. Do you think that would be possible?

[sunxd3]

It makes perfect sense!

In a way AbstractPPL.evaluate!!(model::BUGSModel)does what you suggested, so it is possible to first initialize!(model, param_vector) and then evaluate using the single argument evaluate!!, but like you said, it doesn't a tempered option yet

It's probably time for a refactor now, let me think about it

[miguelbiron]

Awesome!