Thanks to visit codestin.com
Credit goes to github.com

Skip to content

GuoQiushan/MCL

Repository files navigation

Introduction

This is the implematation of Multi-Level Contrastive Learning for Dense Prediction Task (MCL). Our code is based on Openselfsup. For the installation and data preparation, please refer to the INSTALL.md.

avatar

Pre-train

Our codebase supports distributed training. All outputs (log files and checkpoints) will be saved to the working directory, which is specified by work_dir in the config file.

./tool/dist_train.sh configs/selfsup/MCL/MCL.py 8

The default learning rate in config files is for 32 GPUs and 128 images (batch size = 32*128 = 4096).

The released pre-trained model weights can be found here

Finetune

  1. Install Detectron2
  2. Convert the pretrained weights:
cd benchmarks/detection;
python convert-sppretrain-to-detectron2.py PATH_TO_PRETRAIN_WEIGHTS.pth PATH_TO_SAVE_D2_WEIGHT.pkl;
python train_net.py \
    --config-file ./configs/coco_R_50_FPN_CONV_1x_moco.yaml --num-gpus 8 \
    --dist-url 'tcp://localhost:10001' MODEL.WEIGHTS PATH_TO_SAVE_D2_WEIGHT.pkl

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published