evaluation
Routes PyTorch Metric Learning questions about AccuracyCalculator, testers, and nearest-neighbor inference.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- evaluation
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Routes PyTorch Metric Learning questions about AccuracyCalculator, testers, and nearest-neighbor inference.
- disable-model-invocation
- true
02 资源文件 3
- 参考文档references/evaluation-and-inference.md4.6 KB
- 参考文档references/troubleshooting.md2.0 KB
- 脚本scripts/smoke_evaluation.py2.9 KB
03 SKILL.md 正文
Evaluation and inference
Use this sub-skill when the user wants to score an embedding space, compare query/reference splits, or perform nearest-neighbor lookup after training.
Typical triggers
- "How do I compute precision@1, mAP, R-precision, NMI, or AMI?"
- "How do I use GlobalEmbeddingSpaceTester or the other testers?"
- "How do I do nearest-neighbor search on embeddings?"
- "How do I save or reload the faiss index?"
- "How do I compare query and reference splits or use a custom label-comparison function?"
In scope
- Accuracy:
AccuracyCalculatorand its metric family (precision_at_1,r_precision,mean_average_precision,mean_average_precision_at_r,mean_reciprocal_rank,NMI,AMI). - Testers:
GlobalEmbeddingSpaceTester,WithSameParentLabelTester,GlobalTwoStreamEmbeddingSpaceTester, andBaseTester. - Inference:
InferenceModel,MatchFinder,FaissKNN,FaissKMeans,CustomKNN. - Tiny in-memory datasets for evaluation smoke checks, including
EmbeddingDataset.
Out of scope
- Choosing the loss/miner stack belongs in
components. - Training loop, logging, and checkpointing belong in
training. - Dataset download and sampler construction belong in
dataunless the question is purely about evaluation input shape.
How to use this sub-skill
- Read
references/evaluation-and-inference.mdfor the metric map, tester usage, and inference search patterns. - Run
scripts/smoke_evaluation.pywhen you want a tiny confirmation thatAccuracyCalculator, a tester, andInferenceModelstill work together. - Read
references/troubleshooting.mdwhen the failure mentions faiss,k, custom label comparison, or an untrained index. - If the user is still deciding which loss/miner stack to train, route to
componentsbefore finalizing the evaluation settings.
Common routing decisions
- If the user only needs a metric definition or retrieval score, stay here.
- If the user asks how to choose a validation hook or checkpoint policy, route to
training. - If the user asks for the right dataset split or sampler input before evaluation, route to
data.
Useful public facts
AccuracyCalculatorcan include or exclude metrics, andkcan beNone, a positive integer, or"max_bin_count".ref_includes_query=Trueis the right choice when the query set is part of the reference set.GlobalEmbeddingSpaceTesteris the usual default tester for embedding retrieval tasks.WithSameParentLabelTesterexpects hierarchical labels.GlobalTwoStreamEmbeddingSpaceTesterexpects datasets that yield(anchor, positive, label).InferenceModeldefaults to a faiss-backed k-NN index if you do not overrideknn_func.
Read next
references/evaluation-and-inference.mdfor the metric, tester, and inference reference.references/troubleshooting.mdfor faiss,k, and label-comparison failures.scripts/smoke_evaluation.pyfor a tiny evaluation and inference smoke check.
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).