OpenMMLab Codebases with Deployment Guides in MMDeploy 1.3.1

Updated

MMDeploy 1.3.1 documentation includes dedicated deployment guides for eight OpenMMLab codebases: MMPretrain, MMDetection, MMSegmentation, MMagic, MMOCR, MMPose, MMDetection3d, MMRotate, and MMAction2. Each guide sits under the "OpenMMLab Codebase Support" section of the docs and follows the same general workflow — install MMDeploy, convert the model, run inference, and optionally evaluate — but the model configs, checkpoint names, and task-specific output formats differ per codebase.

Where Each Codebase's Guide Lives

The documentation groups codebase guides under OpenMMLab Codebase Support, with one page per codebase:

Codebase Guide page
MMPretrain MMPretrain Deployment
MMDetection MMDetection Deployment
MMSegmentation MMSegmentation Deployment
MMagic MMagic Deployment
MMOCR MMOCR Deployment
MMPose MMPose Deployment
MMDetection3d MMDetection3d Deployment
MMRotate MMRotate Deployment
MMAction2 MMAction2 Deployment

The site also has a Chinese version; the language switch is in the lower-left corner of the layout, per the documentation's welcome page.

How the General Workflow Adapts to Each Codebase

Every codebase guide reuses the same pipeline described in the general "Get Started" and "How to convert model" pages, with codebase-specific inputs:

  1. Prerequisites and installation — install MMDeploy and the target codebase following the general installation page, then confirm the backend you plan to use (e.g., ONNX Runtime, TensorRT, ncnn, OpenVINO) is available.
  2. Model conversion — run the deploy.py conversion script with the codebase's config file and checkpoint. The codebase guide supplies the exact config path and checkpoint URL for its models; the general "How to convert model" page explains the script's arguments.
  3. Inference — use the generated SDK or the Python inference API to run the converted model on a sample input. Each guide shows the expected output for its task (detection boxes, segmentation masks, pose keypoints, OCR results, etc.).
  4. Evaluation — run the codebase's test/evaluation command against the converted model to compare accuracy with the original.

What Changes Between Codebases

  • Config and checkpoint paths: each guide names the specific model config (e.g., a detector config for MMDetection, a segmentor config for MMSegmentation) and its download link.
  • Task output format: detection returns bounding boxes and labels; segmentation returns masks; pose returns keypoints; OCR returns text regions; video understanding (MMAction2) returns action labels.
  • Preprocessing: input normalization and resize rules come from the codebase's config, so the guide's inference example reflects that codebase's data pipeline.

Practical Notes

  • The docs list supported backends separately under Backend Support (ncnn, ONNX Runtime, OpenVINO, PPLNN, SNPE, TensorRT, TorchScript, RKNN, TVM, Core ML). Check that your chosen backend appears in the relevant codebase guide's tested configuration before converting.
  • Benchmark results per model and backend are published under Benchmark, including tests on embedded devices and TVM — useful when picking a backend for a specific codebase.
  • If a model uses custom operators, see Custom Ops for the per-backend operator lists before assuming conversion will succeed unchanged.
  • For models not covered by an existing guide, the Developer Guide explains how to support new models and new backends.

Start from the codebase page that matches your model family; it will point you to the exact config, checkpoint, and commands rather than the generic examples on the Get Started page.