Skip to content

Continual learning for instruction following in CerealBar.

License

Notifications You must be signed in to change notification settings

lil-lab/clif_cb

Repository files navigation

Installation

  1. Create the conda environment: conda create -n cb_vin python=3.7
  2. Make sure that git-LFS is installed before cloning the repo.
  3. Clone the repo.
  4. Install the requirements: pip install -r requirements.txt
  5. Unzip the preprocessed data: unzip preprocessed.zip. This will create a new directory preprocessed/ including two subdirectories: (1) examples/ includes a pickle file for each example in the dataset; the filenames are prefixed with which split the example is from. (2) games/ includes a pickle file for each original CerealBar game, which includes the game ID and all the static environment information for that game.
  6. Make a directory to store experiments: mkdir experiments/

You will need to install pytorch separately according to your machine's requirements. We tested using torch 1.6.0 and 1.2.0. See this page for details.

Subdirectories

  • config/ defines configuration files, including for running the program, model parameters, etc. Configurations are dataclasses and stored as JSON files on disk in the json_configs/ file.
  • data/ defines classes and functions for loading and storing data, including instructions and game information.
  • environment/ defines information about the CerealBar environment, including props, positions, cards, etc.
  • evaluation/ is used to evaluate models; defines metrics, evaluation functions, and evaluation loops.
  • experiments/ contains subdirectories for each experiment ran.
  • inference/ is used to run inference on a model.
  • json_configs/ includes configurations for running the program.
  • learning/ is used to train a model.
  • protobuf/ is used to store the protobuf format for communication with the standalone or the web server.
  • model/ defines the model architectures.
  • preprocessed/ contains preprocessed data (instruction examples and games).
  • simulation/ contains code for simulating the CerealBar game, either through the Unity standalone or a python copy.
  • util/ contains various utilities.
  • web_agent/ contains code necessary for running the model in live games in browser.

About

Continual learning for instruction following in CerealBar.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published