---

name: mmpretrain description: "Operate MMPreTrain model-zoo inference, config-driven training and evaluation, dataset/customization, analysis utilities, checkpoint conversion, and serving workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

MMPreTrain repo skill

Use this skill when a task involves MMPreTrain / OpenMMLab Model Pretraining Toolbox as a package or framework: model-zoo discovery, inference, feature extraction, config-driven training/testing, dataset preparation, registry customization, analysis tools, checkpoint conversion, or serving packaging.

First checks

• Read references/package-overview.md for the package surfaces and optional capability map.
• Read references/installation-and-troubleshooting.md when installation, import, ModelHub, optional dependencies, or backend readiness are uncertain.
• Run scripts/check_mmpretrain_env.py for a no-download package smoke check before relying on model-zoo APIs or backend-specific behavior.
• Read references/repo-provenance.md before deciding whether this skill is stale for a different checkout or package version.

Route by task

User task	Read
List available models, choose a checkpoint policy, use get_model, run image classification, retrieval, captioning, VQA, visual grounding, NLVR, or extract features	sub-skills/model-zoo-inference/SKILL.md
Inspect or modify configs, plan mim train mmpretrain / mim test mmpretrain, choose CPU/GPU/distributed/Slurm launch, use resume/AMP/auto-scale-lr/TTA/output dumps, or plan K-fold	sub-skills/training-and-evaluation/SKILL.md
Prepare CustomDataset, ImageNet, or OpenMMLab 2.0 annotations; inspect dataloader fields; build transform pipelines; register custom datasets/models/metrics/hooks/optimizers	sub-skills/datasets-and-customization/SKILL.md
Analyze JSON logs/results, compute metrics/confusion matrices/FLOPs, visualize CAM/t-SNE/schedulers/datasets, publish/convert checkpoints, reparameterize, or package TorchServe artifacts	sub-skills/tools-analysis-and-deployment/SKILL.md

Install baseline

Typical package install:

pip install -U openmim
mim install "mmpretrain>=1.0.0rc8"

Development/package-tool install:

pip install -U openmim
mim install -e .

Multi-modal workflows need the optional dependency group:

mim install "mmpretrain[multimodal]>=1.0.0rc8"

MMPreTrain expects Python 3.7+, PyTorch 1.8+, mmcv>=2.0.0,<2.4.0, and mmengine>=0.8.3,<1.0.0. Select a CUDA-capable PyTorch/MMCV build only when the task requires GPU or distributed behavior.

Safe operating defaults

1. For offline or architecture-only model work, use get_model(..., pretrained=False) or inferencer pretrained=False; inferencer classes otherwise try default weights.
2. For CPU-only train/test planning, use MIM package commands with --gpus 0 and/or CUDA_VISIBLE_DEVICES=-1.
3. Inspect a merged config and dataset fields before long runs; route config work to training-and-evaluation and dataset schema work to datasets-and-customization.
4. Treat checkpoint downloads, dataset downloads, TorchServe startup, Slurm jobs, and GPU training as resource-requiring actions. Confirm resources before launching them.
5. Use bundled helpers and references here; do not depend on the original repository checkout being present.

Common failure routing

• File ... .mim/model-index.yml does not exist, incompatible mmcv / mmengine, NumPy ABI warnings, or CUDA backend mismatch: references/installation-and-troubleshooting.md.
• Invalid model name, checkpoint download/cache error, device mismatch, or missing multimodal extras during inference: sub-skills/model-zoo-inference/references/troubleshooting.md.
• Bad _base_ inheritance, --cfg-options quoting, load_from versus resume, AMP/TTA, or distributed launch hangs: sub-skills/training-and-evaluation/references/troubleshooting.md.
• Bad annotation paths/labels, missing LoadImageFromFile, unreadable images, or registry KeyError: sub-skills/datasets-and-customization/references/troubleshooting.md.
• Plotting/headless errors, optional Grad-CAM/t-SNE deps, checkpoint format mismatch, EMA publishing, or TorchServe service errors: sub-skills/tools-analysis-and-deployment/references/troubleshooting.md.