README.md

Downstream Multi-Label Classification Tasks

Requirements

In addition to the requirements for SS-PRL pre-training, the following packages are used:

• scikit-learn

Fine-Tuning Linear Classifiers

• Fine-tuning on the COCO dataset:

  DATA="./data/coco"      # path to the coco dataset
  WEIGHT="ss-prl.pth.tar" # pre-trained weight
  
  bash scropts/linear_coco.sh ${DATA} ${WEIGHT}
  

• Fine-tuning on the VOC dataset:

  DATA="./data/voc"       # path to the voc dataset
  WEIGHT="ss-prl.pth.tar" # pre-trained weight
  
  bash scropts/linear_voc.sh ${DATA} ${WEIGHT}
  

Semi-Supervised Training

• Semi-Supervised training on the COCO dataset:

  DATA="./data/coco"      # path to the coco dataset
  WEIGHT="ss-prl.pth.tar" # pre-trained weight
  PERC="1"                # fine-tune with 1 or 10 percent of labels
  
  bash scropts/semisup_coco.sh ${DATA} ${WEIGHT} ${PERC}