How facebookresearch/detectron2 works
Detectron2 is a modular PyTorch computer-vision library for detection and segmentation, used through configurable training, inference, evaluation, and deployment workflows. The main path is configuration and dataset preparation into a trainer, model, and evaluation; pretrained model access and export are additional capabilities. The map emphasizes components and connections supported by the README and source index. Some model and data internals were not sampled, so uncertain links are omitted.
Main entry points: defaults.py Default trainer · projects Research projects · meta_arch Model architecture · roi_heads.py ROI heads
Components 24
Training workflow
- Configuration
detectron2/config/config.py - Dataset registry
detectron2/data/datasets - Data loading
detectron2/data/build.py - Default trainer
detectron2/engine/defaults.py - Training loop
detectron2/engine/train_loop.py - Training hooks
detectron2/engine/hooks.py - Metrics and events
detectron2/utils/events.py - Optimizer and schedule
detectron2/solver/build.py
Vision modeling
- Model architecture
detectron2/modeling/meta_arch - Feature backbones
detectron2/modeling/backbone - Proposal generation
detectron2/modeling/proposal_generator - ROI heads
detectron2/modeling/roi_heads/roi_heads.py - Mask prediction
detectron2/modeling/roi_heads/mask_head.py - Vision structures
detectron2/structures/boxes.py
Evaluation and deployment
- COCO evaluation
detectron2/evaluation/coco_evaluation.py - Model checkpoints
detectron2/checkpoint/detection_checkpoint.py - Model zoo
detectron2/model_zoo/model_zoo.py - Model export
detectron2/export
Research extensions
- Research projects
projects - DensePose
projects/DensePose/densepose/modeling/inference.py - Panoptic DeepLab
projects/Panoptic-DeepLab/panoptic_deeplab/panoptic_seg.py - PointRend
projects/PointRend/point_rend/mask_head.py - TensorMask
projects/TensorMask/tensormask/arch.py
Other components
- Researcher or app
Connections 26
- Researcher or app → Configuration: sets up
- Researcher or app → Default trainer: starts training
- Configuration → Default trainer: configures
detectron2/engine/defaults.py - Default trainer → Model architecture: builds
detectron2/engine/defaults.py - Default trainer → Data loading: builds
detectron2/engine/defaults.py - Dataset registry → Data loading: provides datasets
detectron2/data/__init__.py - Default trainer → Training loop: delegates training
detectron2/engine/defaults.py - Default trainer → Optimizer and schedule: builds optimizer
detectron2/engine/defaults.py - Training hooks → Training loop: observes loop
detectron2/engine/hooks.py - Training hooks → Metrics and events: writes metrics
detectron2/engine/hooks.py - Model architecture → Feature backbones: uses features
- Model architecture → Proposal generation: generates proposals
- Model architecture → ROI heads: dispatches regions
- ROI heads → Mask prediction: builds mask head
detectron2/modeling/roi_heads/roi_heads.py - ROI heads → Vision structures: uses structures
detectron2/modeling/roi_heads/roi_heads.py - Default trainer → COCO evaluation: evaluates models
detectron2/engine/defaults.py - COCO evaluation → Vision structures: converts instances
detectron2/evaluation/coco_evaluation.py - Default trainer → Model checkpoints: loads weights
detectron2/engine/defaults.py - Researcher or app → Model zoo: selects pretrained models
- Model zoo → Model checkpoints: loads model weights
detectron2/model_zoo/model_zoo.py - Researcher or app → Model export: exports model
- Researcher or app → Research projects: uses extensions
- Research projects → DensePose: includes
- Research projects → Panoptic DeepLab: includes
projects/Panoptic-DeepLab/panoptic_deeplab/panoptic_seg.py - Research projects → PointRend: includes
projects/PointRend/point_rend/mask_head.py - Research projects → TensorMask: includes
projects/TensorMask/tensormask/arch.py