policy_gradient_outer_update
Apply the score-function outer update for probabilistic bilevel coreset selection without implicit differentiation.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- policy_gradient_outer_update
- description
- Apply the score-function outer update for probabilistic bilevel coreset selection without implicit differentiation.
02 资源文件 2
- 脚本scripts/policy_update.py1.6 KB
- 脚本tests/test_policy_update.py550 B
03 SKILL.md 正文
Policy Gradient Outer Update
Use this skill when a sampled coreset has been trained in an inner loop and an outer validation loss is available. It implements the paper's policy-gradient estimator and then delegates feasibility to capped-simplex projection.
Do not use this skill for backpropagating through the inner optimizer, Hessian inverse approximations, or greedy subset addition. Its contract is specifically the forward-loss score-function update from Equation (6).
Inputs
- Current probabilities
s. - Sampled binary mask
mfrom the same probabilities. - Scalar validation or outer loss.
- Learning rate
eta. - Coreset budget
K.
Outputs
- Feasible updated probabilities.
- Raw pre-projection values.
- Score-function gradient diagnostics.
Workflow
- Import or otherwise use the mask-relaxation score gradient.
- Multiply the score gradient by the observed outer loss.
- Take a descent step on probabilities.
- Project the result with the capped-simplex projection.
- Log diagnostics so recovery can prove the generated skills were exercised.
Validation
Run python tests/test_policy_update.py or validate the skill tree with tests enabled. The tests confirm finite gradients, probability movement, and post-update feasibility.
Limitations
The update quality depends on the caller's inner-training and validation-loss computation. This skill intentionally does not own model training.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).