Toy ML projects.

Remote development

We are currently using a linux based remote environment with all the cuda dependencies installed.

Create a virtual environment:

mkdir ~/code/.venv
virtualenv ~/code/.venv/ml
source ~/code/.venv/ml/bin/activate
pip install -r bazel_requirements.txt  # TODO(pauldb): Update this.

To run tests:

cd ~/code/ml
PYTHONPATH=. pytest board_games/tests 

To run scripts:

PYTHONPATH=. python board_games/ticket2ride/scripts/simulate.py

Bazel

TODO(pauldb): Get rid of this -- using bazel is annoying as fuck.

These instructions assume bazel 8.0.1 is installed.

bazel_requirements.in lists the packages on which we have actual dependencies. The transitive list of dependencies is generated in bazel_requirements.txt with:

pip-compile bazel_requirements.in --output-file=bazel_requirements.txt --allow-unsafe

Run the test for a package with

bazel test //board_games:board_games_test

or individual tests with

bazel test //board_games:board_games_test --test_arg=board_games/tests/ticket2ride/test_policy_helpers.py::test_get_valid_actions

Weights and Biases

Set the WANDB_ENTITY and WANDB_API_KEY environment variables in ~/.zshrc.