DynODE Inference Module

DynODE Inference Module#

This document describes the core inference classes and helper utilities used in the DynODE framework for probabilistic compartmental modeling. These APIs are designed to facilitate model fitting, parameter sampling, and checkpointing of simulation states.

Visual Representation#

        %%{ init: { "theme": "dark", "themeVariables": { "primaryColor": "#bb86fc", "background": "#121212"} } }%%
classDiagram
    %% Abstract base class
    class InferenceProcess {
        <<abstract>>
        +**numpyro_model**: Callable
        +inference_prngkey: Array
        +infer(**kwargs)
        +get_samples(group_by_chain=False, exclude_deterministic=True)
        +to_arviz()
        - _inference_complete: bool
        - _inferer: Optional[MCMC | SVI]
        - _inference_state: Optional[HMCState | SVIRunResult]
        - _inferer_kwargs: Optional[dict]
    }
    class MCMCProcess {
        +num_samples: int
        +num_warmup: int
        +num_chains: int
        +nuts_max_tree_depth: int
        +nuts_init_strategy: Callable
        +nuts_kwargs: dict
    }
    class SVIProcess {
        +num_iterations: int
        +num_samples: int
        +guide_class: Type[AutoContinuous]
        +guide_init_strategy: Callable
        +optimizer: _NumPyroOptim
        +progress_bar: bool
        +guide_kwargs: dict
    }

    %% Inheritance
    InferenceProcess --> MCMCProcess : subclass
    InferenceProcess --> SVIProcess : subclass

Inference Classes (dynode.infer.inference)#

InferenceProcess#

Abstract base class for all inference processes in DynODE. Defines the interface for fitting a numpyro_model to data, retrieving posterior samples, and exporting results to ArviZ for diagnostics and visualization.

Key Methods:

  • infer(**kwargs): Abstract. Fit the model to data.

  • get_samples(group_by_chain=False, exclude_deterministic=True): Abstract. Retrieve posterior samples.

  • to_arviz(): Abstract. Convert results to an arviz.InferenceData object.

MCMCProcess(InferenceProcess)#

Implements inference using Markov Chain Monte Carlo (MCMC) with the NUTS sampler from NumPyro.

Parameters:

  • num_samples, num_warmup, num_chains: Control MCMC sampling.

  • nuts_max_tree_depth, nuts_init_strategy, nuts_kwargs: NUTS sampler configuration.

  • progress_bar: Show progress during sampling.

Key Methods:

  • infer(**kwargs): Runs MCMC and stores the sampler state.

  • get_samples(group_by_chain=False, exclude_deterministic=True): Returns posterior samples, optionally grouped by chain and/or including deterministic sites.

  • to_arviz(): Returns an arviz.InferenceData object with posterior, prior, and posterior predictive samples.

SVIProcess(InferenceProcess)#

Implements inference using Stochastic Variational Inference (SVI) with NumPyro’s autoguides.

Parameters:

  • num_iterations, num_samples: Control SVI fitting and posterior sampling respectively.

  • guide_class, guide_init_strategy, guide_kwargs: Guide configuration.

  • optimizer: SVI optimizer (default: Adam).

  • progress_bar: Show progress during fitting.

Key Methods:

  • infer(**kwargs): Runs SVI and stores the optimizer state.

  • get_samples(exclude_deterministic=True): Returns posterior samples from the variational guide. No chains are used in SVI, so group_by_chain is not applicable.

  • to_arviz(): Returns an arviz.InferenceData object with prior, posterior predictive, and log-likelihood.

Inference Gotchas and Tips#

  • For information on exactly what to put inside of numpyro_model, please refer to the library backend documentation, section on NumPyro. As numpyro sites are the primary mechanism for the solver/optimizer of each inference process to update and sample parameters.

  • in the event that your sampler/optimzer

Sampling and Resolution Utilities (dynode.infer.sample)#

sample_distributions(obj, rng_key=None, _prefix="")

Recursively traverses a data structure, sampling any numpyro.Distribution objects found.

  • Handles nested dicts, lists, and Pydantic models.

  • Site names are constructed using the _prefix argument for traceability.

Returns: A copy of obj with all distributions replaced by samples.

resolve_deterministic(obj, root_params, _prefix="")

Recursively resolves any DeterministicParameter objects in a data structure, replacing them with their computed values based on root_params.

Returns: A copy of obj with all deterministic parameters resolved.

sample_then_resolve(parameters, rng_key=None)

Convenience function that:

  1. Deep-copies parameters so that parallel chains of inference do not interfere with each other.

  2. Samples all distributions

  3. Resolves all deterministic parameters

Returns: A fully concrete, JAX-compatible copy of parameters.

Checkpointing Utilities (dynode.infer.checkpointing)#

checkpoint_compartment_sizes(config, solution, save_final_timesteps=True, compartment_save_dates=[])

Records compartment sizes at specified simulation dates for debugging and analysis.

Parameters:

  • config: The SimulationConfig used for the ODE simulation.

  • solution: The diffrax.Solution object from ODE integration.

  • save_final_timesteps: If True, saves the final value for each compartment.

  • compartment_save_dates: List of datetime.date objects to checkpoint.

Behavior:

  • Uses numpyro.deterministic to record compartment values at requested dates and/or at the final timestep.

See Also#

Contents