GitHub

Results

The fit function returns a ADVIResults or ABCResults object.

ABCResults

The ABC algorithm returns an ABCResults type containing the full history and final results of the inference process. The type has the following fields:

  • MAP::Vector: Maximum a posteriori estimates of the parameters.
  • theta_history::Vector{Matrix}: History of parameter values across all PMC iterations. Each matrix contains the accepted parameters for that iteration with columns being parameters and rows being samples.
  • epsilon_history::Vector: History of acceptance thresholds (epsilon values) used in each iteration.
  • acc_rate_history::Vector: History of acceptance rates achieved in each iteration.
  • weights_history::Vector{Vector}: History of importance weights for accepted samples in each iteration.
  • final_theta::Matrix: Final accepted parameter values from the last iteration.
  • final_weights::Vector: Final importance weights from the last iteration.

You can access these fields directly from the results type:

results = fit(model, param_dict)

# Maximum a posteriori estimates
map_estimates = results.MAP

# Access final parameter values
final_params = results.final_theta

# Get acceptance rates across iterations
acc_rates = results.acc_rate_history

# Get MAP estimates
map_estimates = results.MAP

ADVIResults

The ADVI algorithm returns an ADVIResults type containing the inference results. The type has the following fields:

  • samples::AbstractArray: Matrix of posterior samples drawn after fitting. Each row represents a sample and each column represents a parameter.
  • MAP::AbstractVector: Maximum a posteriori estimates of the parameters.
  • variational_posterior: The fitted variational posterior distribution containing the full inference results. In Turing.jl, this is obtained via q = vi(model, vi_alg).

You can access these fields directly from the results object:

results = fit(model, param_dict)
posterior_samples = results.samples # Access posterior samples
map_estimates = results.MAP # Get MAP estimates
posterior = results.variational_posterior # Full variational posterior