Taskfile.yml

version: '3'

# requirements: Install taskfile, nvidia-container-toolkit and docker.
# Global Variables
vars:
  GT_IMAGE_NAME: roar-attribution-metrics
  DATA_DIR: /data
  PYTHON_FLAGS: "TORCH_HOME=./.pytorchcache/ PYTHONPATH=. "
  DOCKER_FLAGS: "--shm-size 8G --gpus all --network=host --ipc=host -it --rm "

tasks:

  install-nvidia-container-toolkit:
    desc: https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker
    cmds:
      - echo "See https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker if this doesnt work."
      - curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey | sudo apt-key add -
      - distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
      - |
        curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list | \
        sudo tee /etc/apt/sources.list.d/nvidia-container-runtime.list
      - sudo apt-get update
      - sudo systemctl restart docker

  build-docker:
    desc: Build docker image
    cmds:
      - |
        docker build -t {{.GT_IMAGE_NAME}} - < ./Dockerfile

  _run-in-docker:
    cmds:
      - |
        docker run {{.DOCKER_FLAGS}} \
        -v {{.PWD}}/:/workspace/ \
        -v {{.DATA_DIR}}/:/data/ \
        -v $(pwd)/.bash_history:/.bash_history \
        {{.GT_IMAGE_NAME}} \
        {{.GT_CMD}}

  bash-docker:
    desc: Run bash in docker container
    cmds:
      - task: _run-in-docker
        vars:
          GT_CMD: bash

  train-cls:
    desc: Train a classification network on PASCAL dataset. To be run from inside docker. Optional var - CONFIG_PATH
    cmds:
    - |
      PYTHONPATH=. TORCH_HOME=./.py-cache python src/roar/train_model.py

  extract-attribution-maps:
    desc: This task computes attribution maps and extracts them to specified directory.
    cmds:
      - |
        PYTHONPATH=. TORCH_HOME=./.py-cache python src/roar/generate_attribution_dataset.py

  evaluate-attribution-maps-roar:
    desc: This task computes roar metrics on extracted attribution maps.
    cmds:
      - |
        PYTHONPATH=. TORCH_HOME=./.py-cache python src/roar/retrain_and_evaluate.py

  roar:
    desc: Run full roar pipeline.
    cmds:
      - task: train-cls
      - task: extract-attribution-maps
      - task: evaluate-attribution-maps-roar

  lerf:
    desc: This task computes lerf metrics. By default uses Resnet50 and ImageNet dataset.
    cmds:
      - |
        PYTHONPATH=. TORCH_HOME=./.py-cache python src/lerf/pixel_perturbation_analysis.py