joakimjohnander / DGPNet

Dense Gaussian Processes for Few-Shot Segmentation
49 stars 4 forks source link

DGPNet - Dense Gaussian Processes for Few-Shot Segmentation

Welcome to the public repository for DGPNet. The paper is available at arxiv: https://arxiv.org/abs/2110.03674 .

How to run

Download data

  1. Download and unzip PASCAL and COCO images
  2. Download and unzip PASCAL and COCO annotations (we provide link here)
  3. Change local_config.py to point out the images and annotations. Also change slurm_launch.sh if using slurm.
  4. Download and unzip PASCAL and COCO data splits (we provide link here)
  5. Make sure that the data splits are at DGPNet/data_splits

Install dependencies

The dependencies are listed in DGPNet/singularity/Dockerfile21.09

Train and test model

We typically run via slurm, using

sbatch singularity/slurm_launch.sh runfiles/dgp_5shot_pascal_resnet50.py --train --test --dataset pascal --fold 0 --add_packages_to_path

Code layout