Skip to content

Latest commit

 

History

History
76 lines (61 loc) · 3.52 KB

README.md

File metadata and controls

76 lines (61 loc) · 3.52 KB
Raw

Getting started

  1. Clone https://github.com/BASALT-2022-Karlsruhe/ka-basalt-2022-datadownloader somewhere outside of this project. For example to a shared folder, where all on your server have access to:
  2. Download the basalt data
    1. Clone git clone [email protected]:BASALT-2022-Karlsruhe/ka-basalt-2022-datadownloader.git
    2. Move into dircd ka-basalt-2022-datadownloader
    3. Create an .env file and adjust the number of samples you'd like to download (see ReadMe.md)
    4. run.sh --> You should end up with a volume containing the downloaded demonstration data
  3. Move back to this project: ka-basalt-2022
    1. cd ..
    2. cd ka-basalt-2022
  4. Create an .env file and adjust the parameters
    1. NAME=<NAME>_<Goal> # e.g. kulbach_baseline
    2. VERSION='0_0_1' # Version of your experiments
    3. MODELS_ROOT='/home/shared/BASALT/models' # Folder where you expect and save your models.
    4. PORT=9898
    5. PYTHONUNBUFFERED=1
    6. DATA_ROOT=data_wombat
    7. GIT_ACCESS_TOKEN=YOUR_TOKEN_HERE_123

where DATA_ROOT=data_wombat or DATA_ROOT=data

  • data_wombat: Loads data from volume on mounted shared wombat-server folder
  • data: Loads from volume on host server (Bison)

  1. Build container and start docker container

    1. sh run.sh

  2. This starts bin/bash on the container . From here you can now start e.g. train.py to train your agent

  3. To be able to specify different GPUs for e.g. train.py, change the gpu paramter within the docker_compose.yaml (DO NOT COMMIT CHANGES WITHIN THIS FILE!) for the graphics card you'd like to use.

start bash on container

docker exec -it --user root basalt_container_${NAME}_${GOAL} /bin/bash

Start Training process directly + inpect logs of container

In your docker-compose .yaml (or docker-compose.override.yaml), if you change entrypoint to:

entrypoint: "python train.py"

and just start run.sh, then it will start the training process directly and you should be able to see the output via docker logs -f CONAINER_NAME

Submitting Results

[Official Tutorial https://github.com/minerllabs/basalt_2022_competition_submission_template/blob/main/README.md]

  1. Go to https://gitlab.aicrowd.com/, navigate to "Preferences" -> "SSH Keys" and add an ssh-key to your profile.
  2. Create a private repo.
  3. Add it as a remote via git remote add aicrowd [email protected]:<user>/<repo>.git.
  4. Modify the aicrowd.json file. Use "debug": true when testing the submission process.
  5. Open bash in the docker container.
  6. Run git lfs track train/*.weights.
  7. Check if the model-weights (and other large files you want to push) are marked with Git LFS objects to be committed when calling git lfs status.
  8. If the model weights are not tracked correctly, run git lfs migrate info --everything --include="train/*.weights" followed by git add --renormalize . and check again.
  9. Commit.
  10. Push the branch you want to submit via git push aicrowd <branch>
  11. Create a git tag with git tag -am "submission-<version>" submission-<version>
  12. Push the tag with git push aicrowd submission-<version>
  13. Check the status of your submission in the issues section of the repository.

Testing code on Debian without Docker

# Install dependency packages
sudo apt -qq update && xargs -a apt.txt apt -qq install -y --no-install-recommends \
 && rm -rf /var/cache/*

# Create conda environment
conda env create -n basalt -f environment.yml --prune

# Activate environment
conda activate basalt

# Test whether your code works
python <your-script>.py