计划执行与检查节点
executing-plans
按既定实施计划完成工作,保持过程检查和反馈。
privacy-accounting【效率与协作】Routes TensorFlow Privacy users who want privacy budgets, epsilon/delta statements, or the repo's privacy accounting CLIs.
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
Use this sub-skill when the user wants to compute privacy budgets, find a noise multiplier, or explain the accounting assumptions behind a DP-SGD run.
compute_dp_sgd_privacy / compute_dp_sgd_privacy_statementcompute_noise_from_budgetRDP and PLD accountant choicesmax_examples_per_user matters../training/DPQuery internals -> ../queries/../privacy-tests/../fast-clipping/references/api-reference.md for the verified function signatures and the CLI flag map.references/troubleshooting.md for missing-flag, delta, user-level, and accountant-selection failures.../../references/install-and-scope.md for the minimum runtime.RDP unless the user explicitly wants PLD.scripts/compute_dp_sgd_privacy.py mirrors the privacy-statement CLI with the same core parameters.scripts/compute_noise_from_budget.py mirrors the inverse search CLI that solves for a noise multiplier.../training/ usually feeds the values that this sub-skill consumes.../queries/ owns the lower-level query math that eventually feeds some accounting paths.