mir_recovery_evaluation
Run bounded MIR recovery experiments and emit accuracy, forgetting, traces, and mechanism evidence.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- mir_recovery_evaluation
- description
- Run bounded MIR recovery experiments and emit accuracy, forgetting, traces, and mechanism evidence.
02 资源文件 2
- 脚本scripts/recovery_eval.py9.7 KB
- 脚本tests/test_recovery_eval.py1.0 KB
03 SKILL.md 正文
MIR Recovery and Evaluation Harness
Use this skill after MIR module skills exist and a runtime handoff has been produced. It orchestrates a small online continual-learning experiment, invokes or cross-checks the generated memory and scoring skills, and writes recovery artifacts suitable for validation.
Inputs
- Attempt directory containing
module_plan.jsonandenvironment/runtime_handoff.json. - Generated skills root containing online memory, virtual-update interference, and optional latent MIR helpers.
- Output directory under
recovery/. - Fixed seed and reduced/proxy declaration when full benchmark recovery is blocked.
Outputs
recovery_result.jsonwith paper target metadata, numeric metrics, commands, and mechanism checks.logs/training_trace.jsonwith updates, selected replay examples, losses, and parameters.logs/generated_skill_invocations.jsonlisting generated skill usage.source_manifest.jsonlisting allowed sources.
Workflow
- Load
module_plan.json.fast_recovery_targetwithout changing dataset, metric, or target metadata. - Build a deterministic non-iid stream with three small binary tasks.
- Run MIR and a baseline selector using the generated online memory and virtual-update scoring contracts.
- Evaluate final accuracy and forgetting from task-level histories.
- Cross-check the latent MIR helper with a tiny probability-vector fixture when no real generator is available.
- Save all JSON artifacts and report proxy limitations explicitly.
Validation
Run the script directly in a recovery directory, then run the Distiller recovery validator. Included unit tests exercise metric computation and artifact structure without using external packages.
Limitations
This skill provides a reduced proxy harness for bounded soft-mode recovery. It does not claim to reproduce the paper's full MNIST, CIFAR-10, or MiniImagenet multi-seed results.
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).