Configuration¶

ObjDet uses LightningCLI for configuration management. This allows you to configure all aspects of training via YAML files.

Configuration Structure¶

A complete configuration includes:

# Model configuration
model:
  class_path: objdet.models.FasterRCNN
  init_args:
    num_classes: 80
    backbone: resnet50_fpn_v2

# Data configuration
data:
  class_path: objdet.data.COCODataModule
  init_args:
    data_dir: /path/to/coco
    batch_size: 16

# Trainer configuration
trainer:
  max_epochs: 50
  accelerator: cuda
  devices: 1
  precision: 16-mixed

Overriding Config via CLI¶

# Override specific values
objdet fit --config base.yaml \
  --model.init_args.num_classes=10 \
  --data.init_args.batch_size=32

# Multiple config files (merged)
objdet fit --config base.yaml --config experiment.yaml

Available Models¶

Model	Class Path	Notes
Faster R-CNN	`objdet.models.FasterRCNN`	Two-stage detector
RetinaNet	`objdet.models.RetinaNet`	Single-stage, focal loss
YOLOv8	`objdet.models.YOLOv8`	Ultralytics wrapper
YOLOv11	`objdet.models.YOLOv11`	Latest YOLO

Available DataModules¶

Format	Class Path	Annotation
COCO	`objdet.data.COCODataModule`	JSON
VOC	`objdet.data.VOCDataModule`	XML
YOLO	`objdet.data.YOLODataModule`	TXT

Environment Variables¶

Variable	Description	Default
`MLFLOW_TRACKING_URI`	MLflow server URL	None
`MLFLOW_EXPERIMENT_NAME`	Experiment name	objdet
`CELERY_BROKER_URL`	RabbitMQ URL	amqp://localhost