Design proposals/ignored variables

[tdiethe]

Design proposal for "ignored" variables.

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[codecov-io]

Codecov Report

Merging #138 into develop will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff            @@
##           develop     #138   +/-   ##
========================================
  Coverage    87.96%   87.96%           
========================================
  Files           78       78           
  Lines         3955     3955           
  Branches       678      678           
========================================
  Hits          3479     3479           
  Misses         308      308           
  Partials       168      168

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 6ae3fa0...109d05a. Read the comment docs.

[meissnereric]

Is there a reason you can't pass a "targets" list, instead of "ignored"? I think we could then use those targets to only compute log_pdf / draw samples for those variables (and their associated dependencies).

Also it's not really an issue that we define mean-field for all variables in the posterior definition, as long as we ignore unnecessary parts during the computation. This I feel less strongly about as it doesn't change the main API, only the create_gaussian_meanfield helper function.

So in the above, it would look something like:

kwargs = dict(x=x, y0=y, targets=[model.y0])
inference.run(max_iter=max_iter, learning_rate=learning_rate, verbose=False, callback=print_status, **kwargs)

And the inference algorithm would figure out the required subgraph to compute those targets automatically. (Just traversing up the FactorGraph.)

[tdiethe]

Do you mean "targets" here in the sense of classification targets, or inference targets? If it's the former, then I think that's pretty specific to the supervised learning setting. If it's the latter, then actually I think this is quite reasonable in many cases - to be able to specify which variables are observed and which variables you're interested in, and only perform computations on the subgraph that these represent. Note that this is the model that Infer.NET uses (set observed variables, declare which variables you are interested in marginal posteriors for. Of course, sometimes you are truly interested in the full posterior (e.g. when performing MCMC) and this makes less sense.