How wongkinyiu/yolov7 works
YOLOv7 is an object-detection framework for training, evaluating, exporting, and running model variants on images and video. The main workflow prepares labeled data, trains a configured detector, then uses its weights for inference or validation; export and Triton serving provide deployment routes. The graph also shows loss computation and model building as internal layers. README claims guide capabilities and workflow; because several entry-point sources were not sampled, some links rely on documented usage rather than inspected call sites.
Main entry points: train.py Train detector · train_aux.py Auxiliary training · yolo.py YOLO model · test.py Evaluate models
Components 18
Training
- Train detector
train.py - Auxiliary training
train_aux.py - Training configs
cfg/training - Dataset config
data/coco.yaml - Dataset loading
utils/datasets.py - Detection losses
utils/loss.py
Detector
- YOLO model
models/yolo.py - Model layers
models/common.py
Inference
- Run detection
detect.py - Detection utilities
utils/general.py
Deployment
- Export models
export.py - Reparameterize
tools/reparameterization.ipynb - Triton client
deploy/triton-inference-server/client.py
Validation
- Evaluate models
test.py - Detection metrics
utils/metrics.py
Other components
- Engineer
- COCO data
- Triton server
Connections 27
- Engineer → Train detector: starts training
- Engineer → Auxiliary training: starts training
- Engineer → Run detection: provides input
- Engineer → Evaluate models: requests evaluation
- Engineer → Export models: requests export
- Engineer → Reparameterize: runs procedure
- Train detector → Dataset config: reads config
- Auxiliary training → Dataset config: reads config
- Train detector → Training configs: reads model config
- Auxiliary training → Training configs: reads model config
- Train detector → Dataset loading: loads batches
- Auxiliary training → Dataset loading: loads batches
- Dataset loading → COCO data: reads images
- Train detector → YOLO model: constructs model
- Auxiliary training → YOLO model: constructs model
- Train detector → Detection losses: computes loss
- Auxiliary training → Detection losses: computes loss
- YOLO model → Model layers: uses layers
- Train detector → YOLO model: saves weights
- Auxiliary training → YOLO model: saves weights
- Run detection → YOLO model: runs weights
- Run detection → Detection utilities: processes predictions
- Evaluate models → YOLO model: evaluates weights
- Evaluate models → Dataset loading: loads validation data
- Evaluate models → Detection metrics: measures results
- Export models → YOLO model: exports model
- Triton client → Triton server: sends inference