How shiqiyu/opengait works
Project Type & Purpose
• This is a full-stack machine-learning research codebase (“OpenGait”) for gait recognition.
• It orchestrates data preparation (many datasets + pretreatment scripts), model configuration (YAML files under configs/), training & evaluation pipelines (opengait/), and logging/checkpointing (output/, train.sh, test.sh).
File-Structure Analysis
• Top-level directories:
– configs/: YAML configurations for dozens of gait-recognition models and benchmarks.
– datasets/: Raw dataset definitions (JSON) + preprocessing scripts.
– opengait/: Core Python package – data loaders, transforms, models, losses, evaluators, main entrypoint.
– docs/: User tutorials & model-zoo results.
– misc/: Utility shell/Python scripts for cleaning and downloading.
– assets/: Images and logos referenced in README.
– output/: Checkpoint & tensorboard log stubs.
• Scripts: train.sh/test.sh launch opengait/main.py with selected config.
• CI: .github/workflows for stale-issue management.
Key Architectural Patterns
• Modular layering – clear separation of Data Layer, Model Layer, Training/Evaluation Orchestrator, Configuration Layer.
• Factory/Registry Pattern – model classes and datasets are registered and instantiated via config.
• PyTorch DDP & AMP support integrated into training pipeline.
Main Components & Interactions
[CONFIG MANAGER]
• Loads YAML (configs/) → instantiates Dataset, DataLoader, Model, Loss, Optimizer, DDP/AMP flags.
[DATA INGESTION]
• Raw datasets (datasets/) + pretreatment scripts → unified Dataset classes (opengait/data/dataset.py)
• Transform & sampler modules (opengait/data/transform.py, sampler.py, collate_fn.py)
• DataLoader (PyTorch) → batch tensors
[MODEL & LOSSES]
• Backbones (GCN, ResNet, ViT modules in opengait/modeling/backbones/)
• Head & full models (BigGait, DeepGaitV2, GaitBase, etc. in opengait/modeling/models/)
• Loss functions (ce, triplet, supcon in opengait/modeling/losses/)
• LossAggregator combines multi-losses
[TRAINING & EVALUATION]
• main.py orchestrates train/eval loops, handles DDP setup, checkpointing, AMP
• Evaluator & metrics (opengait/evaluation/) → compute accuracy, re-ranking, save results
[OUTPUT & LOGGING]
• Checkpoints & logs written to output/
• TensorBoard logging via opengait/utils/msg_manager.py