peal.entrypoints.train_distilled_predictor

Distil a trained predictor into a smoother surrogate for counterfactuals.

Drives distill_predictor in peal/training/trainers.py: the predictor of --config_predictor is loaded together with its training and validation dataloaders, and a second network described by --config_distilled_predictor is fitted to mimic it with leakysoftplus activations in place of ReLU, so that gradients through it are informative. The explainers perform the same step on the fly (the distilled_predictor field of the explainer config); this script runs it in isolation.

Invocation:

python train_distilled_predictor.py --config_predictor <predictor.yaml> \
    --config_distilled_predictor <distilled_predictor.yaml>

Both arguments are required, so the sys.argv[-1] fallback and the commented-out ModelTrainer block are dead code. A <PEAL_BASE> prefix in the distilled config’s model_path is resolved to the project directory; the distilled model (model.cpl and config.yaml) is written there.

Functions

main	Load both predictor configs and distil the first into the second.