LILO: Learning Interpretable Libraries by Compressing and Documenting Code (arxiv/2310.19791)
To run LILO, you'll need an environment with:
- Python 3.7 (for backwards compatibility with DreamCoder)
- OCaml (required by DreamCoder)
- Rust (required by Stitch)
The easiest way to get this environment is to use the Docker container provided in this repository. If you'd like to build a local environment, see the instructions below.
First, install Docker on your local machine. Then, pull the Docker container from Docker Hub:
docker pull gabegrand/lilo
Alternatively, you can clone the repo and build the Docker container locally:
git clone --recurse-submodules https://github.com/gabegrand/lilo
cd lilo
docker build -t lilo .
LILO is compatible with both Linux (tested on Debian 12 "bookworm" and Ubuntu 20.04 "Focal Fossa") and MacOS (tested on M1 Mac running MacOS 12 Monterey).
- Conda / Miniconda: https://docs.conda.io/en/latest/miniconda.html
- OCaml: https://ocaml.org/install
- Rust: https://www.rust-lang.org/tools/install
To build a local environment, first clone the repo:
git clone --recurse-submodules https://github.com/gabegrand/lilo
cd lilo
Next, install the lilo conda environment.
conda env create -f environment.yml
Note
If this step is taking a long time, you may wish to install the libmamba solver, which significantly speeds up environment resolution.
cd dreamcoder
make setup-ocaml
make
If the environment is installed correctly, you should be able to activate it:
conda activate lilo
Next, from a Python shell, try constructing a simple Program object and performing type inference with .infer() and evaluation via .evaluate():
>>> from run_experiment import *
>>> p = Program.parse("(_rconcat _x _y)")
>>> p.infer()
tsubstr
>>> p.evaluate([])
'xy'
The general entry point for running experiments is run_iterative_experiment.py.
Note
To run experiments with the OpenAI API, you'll need to set the OPENAI_API_KEY environment variable to your API key. You can find your API key at https://platform.openai.com/account/api-keys.
export OPENAI_API_KEY=<sk-...a123>
Each model type is specified in a template file. For example, the template for the lilo model is experiments_iterative/templates/template_lilo.json. The run_iterative_experiment.py script takes an --experiment_type argument that specifies which template to use.
Below, we provide demo commands for running experiments with each model type on the REGEX domain. Note that these commands are designed to be runnable on a consumer-scale machine (e.g. a laptop). See the section below on full-scale experiments for replicating our experiments on a high-performance computing cluster.
experiments_iterative/templates/template_dreamcoder.json
python run_iterative_experiment.py \
--experiment_name test_runs \
--experiment_type dreamcoder \
--domain re2 \
--encoder re2 \
--iterations 1 \
--global_batch_sizes 32 \
--enumeration_timeout 5 \
--recognition_train_steps 100 \
--verbose \
experiments_iterative/templates/llm_solver.json
python run_iterative_experiment.py \
--experiment_name test_runs \
--experiment_type llm_solver \
--domain re2 \
--iterations 1 \
--global_batch_sizes 32 \
--random_seeds 111 \
--init_frontiers_from_checkpoint \
--resume_checkpoint_directory experiments_iterative/outputs/test_runs/domains/re2/dreamcoder/seed_111/dreamcoder_32 \
--verbose \
Important
Requires running python run_iterative_experiment.py with --experiment_type dreamcoder for at least one iteration to generate an initial frontiers.json file.
experiments_iterative/templates/lilo.json
python run_iterative_experiment.py \
--experiment_name test_runs \
--experiment_type lilo \
--domain re2 \
--encoder re2 \
--iterations 1 \
--global_batch_sizes 32 \
--enumeration_timeout 5 \
--recognition_train_steps 100 \
--random_seeds 111 \
--init_frontiers_from_checkpoint \
--resume_checkpoint_directory experiments_iterative/outputs/test_runs/domains/re2/dreamcoder/seed_111/dreamcoder_32 \
--verbose \
Important
Requires running python run_iterative_experiment.py with --experiment_type dreamcoder for at least one iteration to generate an initial frontiers.json file.
There are two main use cases for resuming from a checkpoint:
- Resuming a run that was interrupted (e.g. due to a crash or a timeout)
- Initializing a new run with the frontiers from a prior run or model. This is used above for the LLM Solver and LILO models to provide an initial seed set of programs for few-shot prompting.
To resume from a prior run, use the --resume_checkpoint_directory flag:
--resume_checkpoint_directory experiments_iterative/outputs/<experiment_name>/domains/<domain>/<experiment_type>/seed_<seed>/<experiment_name>_<batch_size>
Note that you will also need to pass --init_frontiers_from_checkpoint to load the frontiers.json file, which contains the set of program solutions from the checkpoint.
By default, resuming from checkpoint will only occur at the first iteration. This is useful in the case where you want to load from an initial frontiers.json and then continue learning from there. To resume from checkpoint at every iteration, use the --init_frontiers_every_iteration flag. This is usually only used for debugging or to resume a run that was interrupted.
The run_iterative_experiment.py script takes a --domain argument that specifies which domain to use. The following domains are supported:
--domain re2 \
--encoder re2 \
--enumeration_timeout 1000 \
--iterations 16 \
--domain clevr \
--encoder clevr \
--enumeration_timeout 600 \
--iterations 10 \
--domain logo \
--encoder LOGO \
--enumeration_timeout 1800 \
--iterations 10 \
Below, we include flags for replicating our experiments on a high-performance computing cluster. We ran our experiments on c5.24xlarge machines on AWS with 96 CPUs.
--global_batch_sizes 96 \
--recognition_train_steps 10000 \
--random_seeds 111 222 333 \
We make all data from our experiments public: s3://lilo-experiments-release/lilo_arxiv_v1/
To view this data in your browser, you can use the AWS CLI (requires AWS account sign-up): https://s3.console.aws.amazon.com/s3/buckets/lilo-experiments-release
@article{grand2023lilo,
title={{LILO}: Learning Interpretable Libraries by Compressing and Documenting Code},
author={Gabriel Grand and Lionel Wong and Matthew Bowers and Theo X. Olausson and Muxin Liu and Joshua B. Tenenbaum and Jacob Andreas},
journal={arXiv preprint arXiv:2310.19791},
year={2023}
}
LILO inherits from the LAPS codebase (github.com/CatherineWong/laps). We gratefully acknowledge Lionel Wong for the use of their codebase as a foundation and for technical support.
MIT License Copyright (c) 2023 Gabriel Grand
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the “Software”), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.