AGENT SKILL

intern-image

Guides InternImage vision foundation-model workflows across image classification, object detection, semantic segmentation, autonomous driving, DCNv3 operators, and TensorRT deployment.

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

01 前言 Front Matter

name
intern-image
license
MIT
metadata
{ "disco-role": "operating" }
description
Guides InternImage vision foundation-model workflows across image classification, object detection, semantic segmentation, autonomous driving, DCNv3 operators, and TensorRT deployment.
disable-model-invocation
true

02 资源文件 5

03 SKILL.md 正文

InternImage Repo Skill

Use this skill when a task mentions InternImage, OpenGVLab InternImage, DCNv3-based vision foundation models, InternImage classification/detection/segmentation configs, InternImage autonomous-driving baselines, or InternImage TensorRT/mmdeploy export.

InternImage is a large-scale vision backbone family built around DCNv3 deformable convolution. The repository is not a single installable Python package: it is a multi-workflow research codebase with separate classification, MMDetection, MMSegmentation, autonomous-driving, and deployment/operator surfaces. Route aggressively before giving commands.

First steps

  1. Identify the user's workflow family, dataset, model size, config/checkpoint, execution backend, and whether they want a command plan, debugging help, or an actual run.
  2. Read references/model-and-config-overview.md when the model family, config name, dataset, or checkpoint pairing is unclear.
  3. Use scripts/check_internimage_environment.py for a safe dependency/backend diagnostic before any GPU-heavy run.
  4. Route to the nearest sub-skill and use that sub-skill's bundled command builder instead of copying long source launch snippets by hand.
  5. For CUDA, DCNv3, mmdeploy, or TensorRT blockers, route to sub-skills/deployment/ even when the original task is classification, detection, or segmentation.

Route map

User task signal Read this next Use when
ImageNet, iNaturalist, classification config YAML, main.py, DeepSpeed, Accelerate, feature extraction, Hugging Face OpenGVLab/internimage_* sub-skills/classification/SKILL.md The task is image classification, backbone hidden states, classification training/evaluation, or Transformers usage.
COCO, LVIS, OpenImages, VOC, CrowdHuman, MMDetection, Mask R-CNN, Cascade, DINO, bbox/mask mAP, detection image demo, SAM-prompted masks sub-skills/detection/SKILL.md The task is object detection or instance segmentation with the MMDetection 2.x InternImage stack.
ADE20K, Cityscapes, COCO-Stuff, Mapillary, NYU Depth V2, Pascal Context, MMSegmentation, UperNet, SegFormer, Mask2Former, mIoU, segmentation image demo sub-skills/segmentation/SKILL.md The task is semantic segmentation with the MMSegmentation 0.x InternImage stack.
Occupancy prediction, BEVFormerOcc, Online HD Map Construction, VectorMapNet, OpenLane-V2, topology metrics, lane centerlines, traffic elements, OpenLane submission validation sub-skills/autonomous-driving/SKILL.md The task is an InternImage autonomous-driving challenge baseline or OpenLane-V2 devkit/schema/metric workflow.
DCNv3, Cuda is not availabel, nvcc, CUDA toolkit, custom operators, ONNX, TensorRT, mmdeploy, export, deployment sub-skills/deployment/SKILL.md The task is backend readiness, operator build, ONNX/TensorRT export, or cross-workflow deployment.

Install and backend posture

  • The repository's documented runtime stacks are older OpenMMLab-era stacks. Avoid silently upgrading to MMDetection 3.x, MMSegmentation 1.x/2.x, or NumPy 2.x unless the user is intentionally porting the code.
  • Common pins from repo evidence include PyTorch 1.10-1.12 with CUDA, timm==0.6.11, mmcv-full==1.5.x, mmdet==2.28.1, mmsegmentation==0.27.0 or 0.29.1 for autonomous OpenLane, numpy<2, pydantic==1.10.13, and yapf==0.40.1 for detection config formatting.
  • Full model execution normally requires CUDA and the DCNv3 operator. A CPU-only command-builder check is not proof of GPU/model runtime.
  • Source DCNv3 builds require a PyTorch CUDA wheel, visible CUDA devices, CUDA_HOME, and nvcc/toolkit compatibility. A GPU driver alone is not enough.
  • TensorRT export additionally requires mmdeploy, TensorRT, CUDNN, and the DCNv3 TensorRT custom backend operator.

Run a safe environment summary from the root skill directory:

python scripts/check_internimage_environment.py --profile detection --profile deployment

For machine-readable output:

python scripts/check_internimage_environment.py --profile autonomous --json

Runtime guardrails

  • Do not claim that this generated skill verified full training/evaluation/export. The verified scope covered static self-containment, helper parser/dry-run checks, and a CPU OpenLane-V2 schema-validation environment; model-scale GPU work remains a user-approved runtime action.
  • Do not tell users to reopen this skill's source evidence. If a command is needed, use the bundled helper in the relevant sub-skill and fill in the user's checkout, data root, checkpoints, and output paths.
  • Keep checkpoint/config/data families aligned. A classification checkpoint is not a full detector/segmentor checkpoint unless the selected config explicitly uses it as a pretrained backbone.
  • Treat network downloads, challenge datasets, pretrained weights, cluster launches, TensorRT builds, and long training as explicit user-controlled actions.

Cross-cutting references

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