How peterande/d-fine works
D-FINE is a real-time object-detection system centered on fine-grained distribution refinement and localization self-distillation. Its principal workflow configures a detector, loads and transforms labeled images, trains and validates the model, then produces postprocessed detections for inference or export. The graph includes documented capabilities whose implementations are not fully sampled; their connections are omitted or marked unverified rather than inferred. Source excerpts show parts of data loading, the decoder, and training, while the README supplies the high-level product and capability context.
Main entry points: det_engine.py Detection engine · inference Inference tools · dfine_decoder.py D-FINE decoder · train.py Training entry
Components 21
Configuration
- YAML config
src/core/yaml_config.py - Component registry
src/core/workspace.py
Data pipeline
- Detection datasets
src/data/dataset/coco_dataset.py - Image transforms
src/data/transforms - Batch loader
src/data/dataloader.py
Detection model
- Feature backbone
src/nn/backbone/hgnetv2.py - Hybrid encoder
src/zoo/dfine/hybrid_encoder.py - D-FINE decoder
src/zoo/dfine/dfine_decoder.py - Denoising queries
src/zoo/dfine/denoising.py - Detection losses
src/zoo/dfine/dfine_criterion.py - Detection postprocess
src/zoo/dfine/postprocessor.py
Training and evaluation
- Training entry
train.py - Detection engine
src/solver/det_engine.py - Training solver
src/solver/det_solver.py - Model validation
src/solver/validator.py - Optimization tools
src/optim/optim.py - EMA weights
src/optim/ema.py
Inference and deployment
- Model export
tools/deployment/export_onnx.py - Inference tools
tools/inference
Other components
- Researcher
- External runtime
Connections 22
- Researcher → Training entry: starts training
- Training entry → YAML config: loads config
train.py - Training entry → Training solver: starts solver
train.py - YAML config → Component registry: builds components
- Batch loader → Detection datasets: reads samples
src/data/dataloader.py - Detection datasets → Image transforms: applies transforms
- Training solver → Detection engine: runs epochs
- Detection engine → Batch loader: iterates batches
src/solver/det_engine.py - Detection engine → Model validation: validates model
src/solver/det_engine.py - Detection engine → Optimization tools: updates weights
- Detection engine → EMA weights: updates average
- Detection engine → Detection losses: computes losses
- Detection engine → Feature backbone: runs detector
- Feature backbone → Hybrid encoder: provides features
- Hybrid encoder → D-FINE decoder: passes features
- D-FINE decoder → Denoising queries: uses queries
src/zoo/dfine/dfine_decoder.py - D-FINE decoder → Detection losses: provides predictions
- Researcher → Model export: requests export
- Model export → Inference tools: exports model
- Researcher → Inference tools: runs inference
- Inference tools → Detection postprocess: formats detections
- Inference tools → External runtime: targets runtime