Separate Sampler and Results classes

[stephengreen]

This splits off importance sampling and related functionality from the Sampler class into the Result class (previously SamplesDataset). The purpose is to simplify and streamline the code, to make it easier to maintain and to write a new inference interface.

Key changes / methods in each class:

• core.result.Result(DingoDataset)
  • Saved attributes include samples, context, event_metadata, log_evidence, and log_evidence_std.
  • importance_sample()
  • subset(parameters): Returns a new Result with a subset of parameters (and dropping weights, log probs, etc.). This is called before training an unconditional flow.
  • train_unconditional_flow(parameters, nde_settings): Trains an unconditional flow model and returns it.
• gw.result.Result(core.result.Result)
  • sample_synthetic_phase()
  • Defines also the GW domain, prior, and likelihood.
• Sampler and subclasses
  • run_sampler(num_samples)
  • log_prob(samples): I still have to check whether this works, but I'm not sure it's really needed. I think it can wait until later.
  • unconditional_model attribute specifies whether the model is unconditional, and if so sampling ignores any event context. This information is stored and accessed from the model metadata. UnconditionalSampler classes were removed.
• GNPESampler
  • No longer uses the use_gnpe_proxy_sampler flag. Instead, when init_sampler gives proxy variables, these are recognized by the GNPE transform and used directly rather than resampled from the kernel.
  • For sampling from the joint distribution $p(\theta, \hat \theta | d, S_n)$ starting from a 2-/3-D unconditional flow, set num_iterations = 1, and then the log_prob will be included in the final result.
• GW Samplers no longer define prior or likelihood, although they still define a domain, which is needed for certain transforms.
• PosteriorModel
  • context and event_metadata are now attributes. I found that it was more straightforward to store this information directly rather than deep within a hierarchy of metadata.

Action required

For detector-time GNPE, I introduced new parameter names L1_time_proxy_relative, etc., to denote the L1_time_proxy after having subtracted off the relative time shifts. This is needed in order to separately preserve L1_time_proxy for training unconditional flows, etc. (This replaces GNPE:L1_time_proxy, etc.) Consequently the GNPE networks are conditioned on L1_time_proxy_relative rather than L1_time_proxy.

To make existing networks compatible with the new GNPEDetectorTimes class, it is necessary to rename the context variables in the .pt files. This can be done using the compatibility script, update_saved_model_gnpe_context_names.py.

Still to-do

• Fix up the docstrings.
• Put the IterationTracker back into GNPESampler._run_sampler().
• Check that log_prob calculations work, but again, I'm not sure that they are needed, since log_prob will usually come along with sampling.

[stephengreen]

I've made various additional improvements to the code:

• Set up the IterationTracker and remove_init_outliers options again for the GNPESampler.
• Saved the importance sampling information, namely the keyword arguments for synthetic phase, and phase and time marginalization. This is stored in a new attribute Result.importance_sampling_metadata. We should also use this to store calibration information and updates to priors, data conditioning, etc.
• Fixed several bugs.