How open-mmlab/mmpretrain works
Project Type and Purpose
• This is a full‐fledged Python toolbox (library/framework) for deep learning pre‐training and multi‐modal tasks, built on PyTorch. It is not a standalone service but a pluggable SDK providing APIs, command‐line scripts, configs, and tools for dataset preparation, training, inference, and evaluation.
High‐Level Directory Analysis
• Top‐level:
• mmpretrain/ → core Python package (APIs, configs, datasets, engine, evaluation, models, utilities, visualization)
• configs/ → model/dataset/training configuration files (YAML/Python) for many backbones and algorithms
• tools/ → CLI scripts (train.py, test.py), model converters, dataset converters, visualization tools
• demo/, projects/ → example usage (image_demo.py, gradio demos, custom projects)
• docs/ → Sphinx documentation source (en/zh_CN)
• docker/, .circleci/, .github/ → CI/CD and container orchestration
• tests/ → unit/integration tests covering all modules
Core Architectural Patterns
• Plugin/Registry Pattern: modules (datasets, models, hooks, metrics) register themselves at import via a central registry.py.
• Config‐Driven Pipeline: user‐provided Python/YAML configs instantiate data pipelines, models, optimizers, runners.
• Engine‐Runner Pattern: training/inference is orchestrated by engine runners (hooks, schedulers, runners).
Main Components
a. User Interface Layer
• CLI Scripts (tools/train.py, tools/test.py, demo/image_demo.py)
• Programmatic APIs (mmpretrain/apis/*.py)
• Configuration Files (configs/*)
b. Core Framework
• Config Parser & Builder (mmpretrain/registry.py, mmpretrain/apis/utils.py)
• Dataset Management (mmpretrain/datasets/*)
• Model Zoo & Builders
• Backbones (mmpretrain/models/backbones/*)
• Heads, Necks, Retrievers, Losses, etc.
• Training Engine (mmpretrain/engine/*)
• Runners, Hooks, Optimizers, Schedulers
• Evaluation Engine (mmpretrain/evaluation/*)
• Metric computations, visual grounding, retrieval metrics
• Utilities & Visualization (mmpretrain/utils/*, mmpretrain/visualization/*)
c. External Dependencies
• PyTorch, torchvision, CUDA, mmcv, mmengine
• Docker container for serving (docker/serve)
• CI/CD (CircleCI, GitHub Actions)
d. Outputs & Artifacts
• Checkpoints, logs, TensorBoard, model zoo entries
• Evaluation reports, visualizations
Essential Elements
• Labels on each component (exact package or directory name)
• Registry as central connector for plugin modules
• Config files → Core pipeline instantiation
• Data → Model → Engine → Evaluation
• Optional: Docker serving & CI/CD pipelines on the side