How open-mmlab/mmaction works
MMAction is a PyTorch toolbox for action understanding. Its user workflow is to prepare video data, select a configured recognition or detection model, train or run inference, and evaluate predictions. The product spans trimmed-video recognition, temporal localization, and spatiotemporal detection, with shared datasets, model components, and bounding-box utilities. The graph preserves those documented capabilities; several subsystem links are documented only at a high level because their implementation files were not sampled, so those links are marked unverified.
Main entry points: train.py Training API · datasets Video datasets · eval_localize_results.py Localization evaluator · TSN3D.py TSN3D
Components 23
Video data
- Data preparation
data_tools - Video datasets
mmaction/datasets - SSN proposals
mmaction/datasets/ssn_dataset.py - Video transforms
mmaction/datasets/transforms.py - Batch loading
mmaction/datasets/loader/build_loader.py
Model components
- Model registry
mmaction/models/registry.py - Model builder
mmaction/models/builder.py - Action recognizers
mmaction/models/recognizers - TSN3D
mmaction/models/recognizers/TSN3D.py - Video backbones
mmaction/models/tenons/backbones - BNInception
mmaction/models/tenons/backbones/bninception.py - Spatial detectors
mmaction/models/detectors - Temporal localizers
mmaction/models/localizers - Training losses
mmaction/losses
Training and inference
- Training entrypoints
tools/train_recognizer.py - Training API
mmaction/apis/train.py
Detection and evaluation
- Box targets
mmaction/core/bbox2d - Anchor targets
mmaction/core/anchor2d/anchor_target.py - Validation hooks
mmaction/core/evaluation/eval_hooks.py - Localization metrics
mmaction/core/evaluation/localize_utils.py - Localization evaluator
tools/eval_localize_results.py
Other components
- Researcher
- Model configuration
configs
Connections 23
- Researcher → Data preparation: prepares data
- Data preparation → Video datasets: feeds datasets
- Researcher → Training entrypoints: starts training
- Model configuration → Training API: configures run
- Training entrypoints → Training API: invokes
- Training API → Batch loading: builds batches
mmaction/apis/train.py - Training API → Video datasets: trains on
mmaction/apis/train.py - Training API → Action recognizers: forwards batches
mmaction/apis/train.py - Model builder → Model registry: looks up modules
- TSN3D → Model builder: builds backbone
mmaction/models/recognizers/TSN3D.py - TSN3D → Video backbones: extracts features
mmaction/models/recognizers/TSN3D.py - BNInception → Model registry: registers backbone
mmaction/models/tenons/backbones/bninception.py - Action recognizers → Researcher: predicts actions
- Spatial detectors → Box targets: uses box utilities
- Anchor targets → Box targets: assigns box targets
mmaction/core/anchor2d/anchor_target.py - Temporal localizers → SSN proposals: localizes proposals
- SSN proposals → Video transforms: uses transforms
mmaction/datasets/ssn_dataset.py - Training API → Training losses: optimizes outputs
- Training API → Validation hooks: registers validation
mmaction/apis/train.py - Localization evaluator → Video datasets: builds test dataset
tools/eval_localize_results.py - Localization evaluator → Localization metrics: evaluates detections
tools/eval_localize_results.py - Validation hooks → Video datasets: reads annotations
mmaction/core/evaluation/eval_hooks.py - Localization evaluator → Researcher: reports metrics