peal.entrypoints.run_cfkd¶
Run Counterfactual Knowledge Distillation (CFKD) from a YAML config.
Drives the CFKD adaptor in
peal/adaptors/counterfactual_knowledge_distillation.py: load the student
predictor, explain it with counterfactuals from the configured generator and
explainer, collect feedback (oracle, human or SAE based) on the
counterfactuals and finetune the student on them, for finetune_iterations
rounds.
Invocation:
python run_cfkd.py --config <adaptor.yaml> [--<any CFKDConfig field> value]
Every field of CFKDConfig is exposed as an optional --flag that
overrides the YAML value (add_class_arguments / integrate_arguments).
Set PEAL_CUDA_SYNC=1 to make CUDA launches synchronous while debugging a
kernel fault.
Reads the adaptor config and the predictor, data, generator and explainer
configs it references. Writes into base_dir: config.yaml, logs/
(TensorBoard), platform.txt, one <iteration>/ directory per round with
the explainer collages and the train_dataset / validation_dataset of
counterfactuals, and the finetuned student as model.cpl.
Functions
Load the |