Propagate Yourself

• Paper : https://arixv.org/abs/2011.10043
• Pytorch 1.7.0, cuda 10.1
• Used 4 GPUS (V100) for training.

TODO:

• Fix pixcontrast modules

Environment Settings

• CLI

  git clone https://github.com/Sungman-Cho/Propagate-Yourself-Pytorch.git
  source install_packages.sh

Unsupervised Training (ImageNet-1K)

• PixPro training

  python train.py --multiprocessing-distributed --batch_size=512 --loss=pixpro

• PixContrast training

  python train.py --multiprocessing-distributed --batch_size=512 --loss=pixcontrast

Transfer learning

before downstream training

• Make your current directory downstream.

• Convert a trained PixPro model to detectron2's format:

  python convert-pretrain-to-detectron2.py '$your_checkpoint.pth.tar' pixpro.pkl

• Convert a trained Pixcontrast model to detectorn2's format:

  python convert-pretrain-to-detectron2.py '$your_checkpoint.pth.tar' pixcontrast.pkl

VOC

• Training Epochs: 24K iter

• Image size : [480,800] in train, 800 at inference.

• Backbone : R50-C4

• Training

  # baseline training
  source train_voc_base.sh
  
  # pixpro training
  source train_voc_pixpro.sh
  
  # pixcontrast training
  source train_voc_pixcontrast.sh

COCO

• Followed 1x settings (detectron2)

• Backbone : R50-C4

• Training

  # baseline training
  source train_coco_base.sh
  
  # pixpro training
  source train_coco_pixpro.sh
  
  # pixcontrast training
  source train_coco_pixcontrast.sh