Bayesian Neural Field models for prediction in large-scale spatiotemporal datasets
93
stars
10
forks
source link
Fix sample_size for VI to be the number of posterior samples not number of Monte Carlo samples for divergence #20
Closed
fsaad closed 9 months ago
The
sample_sizeargument infitfor BayesianNeuralFieldVI https://github.com/google/bayesnf/blob/2debb69cf21daf7de982843a82436a57d5a97dd3/src/bayesnf/spatiotemporal.py#L379is supposed to denote the number of posterior samples from each surrogate in the ensemble. However, tracking down further in
inference.fit_vi, we seesample_size: the number of Monte Carlo samples for estimating divergence.num_samples: (not exposed in the API) the number of posterior samples from each surrogate.https://github.com/google/bayesnf/blob/2debb69cf21daf7de982843a82436a57d5a97dd3/src/bayesnf/inference.py#L653-L655