How kaiyangzhou/deep-person-reid works
Project type and purpose
• Torchreid is an open-source PyTorch-based library for person re-identification.
a. CLI / Orchestrator (scripts/)
• scripts/main.py: entry point.
• scripts/default_config.py: default hyperparameters.
• Makefile / Dockerfile: build & deploy.
b. Configuration layer (configs/)
• YAML files: experiment-specific hyperparams.
• Parsed by a config loader in scripts/main.py.
c. Core Library (torchreid/)
• Data Module (torchreid/data/)
– ImageDataManager / VideoDataManager
– Datasets, Samplers, Transforms
• Model Module (torchreid/models/)
– Backbones (ResNet, OSNet, MobileNet, NASNet, etc.)
– Re-id heads, omni-scale blocks, IBN layers
• Engine Module (torchreid/engine/)
– Base Engine, ImageSoftmaxEngine, TripletEngine, Video variants
– Encapsulates training loop, forward/backward pass, evaluation hooks
• Losses Module (torchreid/losses/)
– CrossEntropy, Triplet, Hard-Mining, Label-Smooth
• Metrics Module (torchreid/metrics/)
– Accuracy, Rank, Distance; Cython extensions for fast ranking
• Optimizers & Schedulers (torchreid/optim/)
– Adam, RAdam, LR schedulers (single-step, Warmup, Cosine)
• Utils (torchreid/utils/)
– FeatureExtractor, Logger, Re-ranking, TensorBoard support
d. Example Projects (projects/)
– DML (Deep Mutual Learning)
– OSNet_AIN (NAS search)
– attribute_recognition (person attribute labels)
– Each contains its own main.py, configs, and custom code
e. Tools (tools/)
– compute_mean_std.py: dataset statistics
– export.py: model export (ONNX, OpenVINO, TFLite)
– parse_test_res.py: aggregate results
– visualize_actmap.py: activation map visualization
f. External Dependencies & Services
• PyTorch + CUDA GPU
• TensorBoard for visualization
• ONNX / OpenVINO / TFLite runtimes (export)
• Standard re-id datasets (Market1501, Duke, MSMT17, etc.)
1. User invokes CLI → parses command-line args + YAML config
2. CLI → Config Loader → populates a config object
3. CLI → DataManager Factory → Data Module → DataLoader → Engine
4. CLI → ModelBuilder Factory → Core Model Module → Engine
5. CLI → OptimizerBuilder & SchedulerBuilder → Engine
6. Engine → Loss Module + Metrics Module
7. Engine → Torch backend (autograd, GPU) → parameter updates
8. Engine → Logger / TensorBoard → persist training curves
9. Engine → Checkpoint Saver → disk (model.pth.tar)
10. CLI / Tools/export.py → loads checkpoint → performs model export
11. Example Projects → reuse Core Library via import torchreid → override or extend components
Architectural patterns & design principles
• Factory / Registry pattern for models, optimizers, schedulers (build_XXX APIs)
• Separation of concerns: data loading, model definition, training loop, evaluation, utilities
• Config-driven experiment management (YAML + defaults)
• Plugin-style extensibility: drop-in new datasets, samplers, losses, metrics
• High-performance extensions: Cython for re-ranking; multi-GPU via DataParallel