detection

Downstream Object Detection Tasks

Requirements

We use detectron2 to train the object detection models. To install the package, please follow the installation instructions.

Data & Model Preparation

• Put the COCO dataset under benchmarks/detection/datasets directory, following the directory structure requried by detectron2.

• Convert the pre-trained backbone weights to detectron2's format:

  cd benchmarks/detection
  
  WEIGHT_FILE="ss-prl.pth.tar" # pre-trained weight
  OUTPUT_FILE="ss-prl.pkl"     # converted output file
  
  python convert-pretrain-to-detectron2.py ${WEIGHT_FILE} ${OUTPUT_FILE}
  

Start training

CONVERTED_WEIGHT="ss-prl.pkl" # converted pre-trained weight
LOG_DIR="./log"               # where to write training logs

bash run.sh ${CONVERTED_WEIGHT} ${LOG_DIR}