CP2: Copy-Paste Contrastive Pretraining for Semantic Segmentation

This repo is the codebase of CP2.

• CP2: Copy-Paste Contrastive Pretraining for Semantic Segmentation, in ECCV 2022.

Pretrained Models

The pretrained CP2 models are available as below:

• CP2 from scratch: cp2-r50-aspp-200ep.pth, cp2-r50-aspp-600ep.pth
• CP2 Quick Tuning (20ep) from MoCo v2 models: cp2-r50-aspp-820ep.pth, cp2-vits16-aspp-320ep.pth

Installation

The implementation is built on top of mmseg and moco. Please install with the following steps:

git clone https://github.com/wangf3014/CP2.git
cd CP2/

# We recommend installing CP2 with torch==1.7.0 and mmcv==1.3.5
# Install mmcv (https://github.com/open-mmlab/mmcv). Make sure the mmcv version matches your torch version.
pip install mmcv-full==1.3.5 -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.7.0/index.html
pip install -r requirements.txt
chmod u+x tools/dist_train.sh

Pretraining

For pretraining CP2 from scratch, run the following command:

python main.py --data PATH_TO_YOUR_IMAGENET \
    --config config/config_pretrain.py \
    --epochs 200 --lr 0.015 -b 256

For Quick Tuning, you should first edit the config file config/config_pretrain.py, setting the pretrained_path to your pretrained backbone, and run the command below:

python main.py --data PATH_TO_YOUR_IMAGENET \
    --config config/config_pretrain.py \
    --epochs 20 --lr 0.015 -b 256

Finetuning

We recommend finetuning on multiple GPUs. For finetuning, you should first specify pretrain_path and data_root in config/config_finetune.py

# Please specify the NUM_GPU and YOUR_WORK_DIR
./tools/dist_train.sh configs/config_finetune.py NUM_GPU --work-dir YOUR_WORK_DIR

If using a part of GPUs on your device (e.g. 4/8), you should run finetuning with the following code:

CUDA_VISIBLE_DEVICES=0,1,2,3 PORT=23333 ./tools/dist_train.sh configs/config_finetune.py NUM_GPU --work-dir YOUR_WORK_DIR

Citation

@article{wang2022cp2,
  title={CP2: Copy-Paste Contrastive Pretraining for Semantic Segmentation},
  author={Wang, Feng and Wang, Huiyu and Wei, Chen and Yuille, Alan and Shen, Wei},
  journal={arXiv preprint arXiv:2203.11709},
  year={2022}
}