training-evaluation
Launch, troubleshoot, and inspect MMSegmentation training, validation, testing, metrics, logs, and distributed jobs.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- training-evaluation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Launch, troubleshoot, and inspect MMSegmentation training, validation, testing, metrics, logs, and distributed jobs.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/distributed-training.md6.9 KB
- 参考文档references/metrics-and-logs.md8.0 KB
- 参考文档references/train-test-workflows.md8.3 KB
- 参考文档references/troubleshooting.md9.5 KB
- 脚本scripts/analyze_mmseg_log.py5.9 KB
- 脚本scripts/mmseg_test_wrapper.py12.7 KB
- 脚本scripts/mmseg_train_wrapper.py11.0 KB
03 SKILL.md 正文
Training and Evaluation
Use this sub-skill when the task is to run or debug MMSegmentation training, validation, testing, metrics, log analysis, checkpoint resume/fine-tuning, distributed launch, Slurm jobs, AMP, TTA, offline result formatting, or test-time visualization.
Route Here When
- The user needs a safe MMSegmentation training/testing command,
Runner.from_cfgguidance, distributed launch, or Slurm launch. - A run failed around checkpoints,
work_dirs,--resume,load_from,--amp,--cfg-options, launcher settings, ports, ranks, or unavailable CUDA. - The user needs IoU, Cityscapes, or depth metric behavior explained, including
output_dir,format_only, and saved prediction files. - The user wants to summarize MMSegmentation JSON logs, compare loss/metric curves, or inspect training progress without starting a long run.
- Testing needs
--out, TTA, visualized predictions, or hidden-test/offline submission formatting.
Do Not Handle Here
- Dataset registration, data roots, annotation format, pipelines, or config authoring details; route those to
data-configuration. - Implementing models, decode heads, losses, optimizers, hooks, or registries beyond runtime troubleshooting; route those to
model-customization. - Inference-only APIs such as
init_model,inference_model, or inferencers; route those toinference.
Start With
- Identify whether the user wants a dry-run command, a real execution, or a diagnosis of an existing log/error.
- For training/testing commands, prefer the bundled safe wrappers first:
scripts/mmseg_train_wrapper.py --helpscripts/mmseg_test_wrapper.py --help
- For log summaries, use
scripts/analyze_mmseg_log.py --help; it has no MMSegmentation import requirement and works on small JSON-line fixtures. - If direct commands are clearer than wrappers, use the exact patterns in
references/train-test-workflows.mdandreferences/distributed-training.md. - For metric/output questions, check
references/metrics-and-logs.mdbefore changing evaluator config. - For failures, match the symptom in
references/troubleshooting.mdbefore changing code or re-running expensive jobs.
Key References
references/train-test-workflows.mdcovers single-process training/testing,Runner.from_cfg, AMP, resume/fine-tune, TTA, visualization, output formatting, andwork_dirs.references/metrics-and-logs.mdcoversIoUMetric,DepthMetric, Cityscapes-style evaluation, JSON log summaries, confusion-matrix inputs, and benchmark cautions.references/distributed-training.mdcoverstorch.distributed.launch, multi-node environment variables, Slurm variables, port conflicts, and NPU launch shape.references/troubleshooting.mdmaps common training/evaluation failures to checks and safe fixes.
Safety Defaults
- The bundled training/testing wrappers are dry-run/preflight tools unless
--executeis passed. - Do not start full training, testing, distributed, Slurm, or benchmark jobs without explicit user approval and a bounded config/checkpoint scope.
- Prefer short help checks, config parsing, tiny log fixtures, or metric unit tests before native long-running jobs.
- Keep offline formatting and hidden-test workflows explicit: set evaluator
format_only=Trueonly when ground truth is unavailable or evaluation should be skipped.
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).