Train Configuration

Source code:

• bff/workflows/configs.py
• bff/workflows/train.py
• bff/bayes/learning.py

Purpose

bff train fits one surrogate model per QoI dataset and writes one .lgp file per QoI.

Minimal Example

log: out.log
datasets:
  rdf:
    data: ../03-qoi/qoi-rdf.pt
    mean: sigmoid
training:
  model_dir: ./models
  reuse_models: true
  committee_size: 1
  device: cuda

Top-Level Keys

• log Workflow log file.
• datasets Non-empty mapping from QoI name to dataset settings.
• training Surrogate-training settings.

datasets.<name> Keys

• data QoI dataset file produced by bff qoi.
• mean Mean-function setting passed into surrogate training.
• nuisance Optional fixed nuisance standard deviation.
• observation_scale Optional scaling applied to the effective observation count in the likelihood.
• model Optional model output path. Defaults to training.model_dir/<name>.lgp.

training Keys

• model_dir Output directory for surrogate model files.
• reuse_models Reuse existing model files if present.
• n_hyper_max Maximum number of hyperparameter optimization steps.
• committee_size Number of committee members for model averaging.
• test_fraction Held-out validation fraction.
• device Torch device, for example cpu or cuda.
• any additional keys Forwarded as optimization keyword arguments.

Outputs

bff train writes one .lgp file per QoI.

Each model file stores:

• the trained committee
• the effective observation count used in the likelihood
• the reference observation vector needed later by bff learn