AGENT SKILL

model-customization

Customize MMSegmentation models, registries, projects, checkpoints, and model-complexity checks.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
model-customization
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Customize MMSegmentation models, registries, projects, checkpoints, and model-complexity checks.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

Model Customization

Use this sub-skill when extending MMSegmentation with custom model components, choosing or adapting model zoo configs, validating registry wiring, converting checkpoints, or explaining model-complexity/deployment caveats.

Scope

This sub-skill covers:

  • MMSegmentation model architecture: segmentors, data preprocessors, backbones, necks, decode heads, auxiliary heads, losses, metrics, optimizers, and schedulers.
  • Registry usage through mmseg.registry and legacy builder aliases exposed from mmseg.models.
  • Custom components implemented in-tree, in project packages, or in external Python packages imported by config.
  • Model zoo/config selection, optional projects, checkpoint conversion planning, model publishing, and FLOPs/parameter caveats.

Use sibling skills instead for:

  • Dataset layout, transforms, palettes, and evaluators beyond metric registration: data-configuration.
  • Launching train/test jobs, distributed runners, and result reporting: training-evaluation.
  • Runtime inference APIs, visualization, or serving: inference.

Fast Routing

  • Need to understand what can be customized or which built-ins exist? Read model-components.
  • Need to add a custom backbone, decode head, loss, metric, optimizer, scheduler, or project package? Read customization-recipes.
  • Need to adapt pretrained weights, publish a checkpoint, or explain why a conversion is unsafe? Read checkpoint-conversion.
  • Need to debug KeyError, custom_imports, default_scope, shape mismatch, optional dependency, FLOPs, or MMCV-op failures? Read troubleshooting.
  • Need to check whether a type is registered after importing optional modules? Run inspect_registry.py.

Default Workflow

  1. Identify the target extension point: MODELS for segmentors/backbones/necks/heads/losses/data preprocessors, METRICS for evaluators, OPTIMIZERS or OPTIM_WRAPPER_CONSTRUCTORS for optimizer behavior, and PARAM_SCHEDULERS for schedulers.
  2. Confirm the component is imported before config build: either import it from the package __init__.py path or add custom_imports = dict(imports=[...], allow_failed_imports=False) to the config.
  3. Keep default_scope = 'mmseg' unless intentionally composing registries from another OpenMMLab package.
  4. Smoke-check registry visibility with scripts/inspect_registry.py before debugging model math.
  5. Build or instantiate the smallest safe config first, then run a tiny forward/loss test before launching training.
  6. Treat checkpoint conversion and FLOPs as verification tasks, not guarantees: both depend on exact source checkpoint schemas, target key names, optional dependencies, and supported operators.

Guardrails

  • Prefer mmseg.registry.MODELS.build(...) for new guidance; legacy build_backbone, build_head, build_loss, and build_segmentor still exist but warn that MODELS.build() is preferred.
  • Register custom model classes with @MODELS.register_module() and custom metrics with @METRICS.register_module().
  • A config type='MyType' only works after Python imports the module that runs the registration decorator.
  • Do not promise that a source checkpoint can be loaded or converted unless key mapping, tensor shapes, class counts, and optional dependency versions are checked.
  • Do not treat FLOPs as paper-ready without verifying unsupported ops, post-processing omissions, input shape, and head support.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351