This readme has info on setting up Dreamcoder and working with our codebase. For more details on using Dreamcoder, you can look at readme inside the ec directory.

Ideally, clone it to openmind…

ssh [username]@openmind7.mit.edu
cd /om2/user/[username]
git clone https://github.com/anshula/neurosymbolic-modules.git

...but if you don’t have access to openmind, clone it to another linux machine. The machine probably needs to be linux in order for singularity to run on it smoothly.

First make sure you are in the neurosymbolic-modules folder, then:

cd ec
git clone https://github.com/insperatum/pinn.git
cd pinn
git checkout 1878ef5

cd ..
git clone https://github.com/insperatum/pregex.git
cd pregex
git checkout b5eab11

cd ..
git clone https://github.com/hans/pyccg.git
cd pyccg
git checkout c465a23

Download the singularity container.img file, and put it directly inside the neurosymbolic modules/ec folder.

• Quick way (works on openmind): copy it from Anshula’s openmind folder: srun cp /cbcl/cbcl01/anshula/shared/container.img /om2/user/$USER/neurosymbolic-modules/ec/container.img
• If you need it off of openmind, can copy it from Simon's dropbox: https://www.dropbox.com/s/as6yewdcm2g7tnc/container.img?dl=0
• Longer way: Install it from source using the instructions here: https://github.com/ellisk42/ec

If using openmind, singularity is already installed, so you just have to import it:

• module add openmind/singularity

If on Linux, install with:

• sudo apt-get install singularity-container
• (And then look at https://sylabs.io/guides/3.4/user-guide/installation.html if that doesn't work)

If on macOS:

• Install vagrant using brew.
• Run vagrant up in the root directory of this repo, this launches a VM according to Vagrantfile. Customize this config file as you see fit.
• Run vagrant ssh from the same directory to access the VM. Singularity will be installed inside the VM. Access the git repo from inside the VM by navigating to /vagrant.

Make sure singularity --version returns a version number to make sure the installation was successful.

singularity exec container.img python -u bin/arc_demo.py -t 100 -g -i 5

You should see some output enumerating solved tasks.

To learn how to solve your own tasks from scratch, follow the tutorial from https://github.com/ellisk42/ec/blob/master/docs/creating-new-domains.md

For more on the ARC-specific infrastructure.

• To change the primitives or tasks used in a given run, edit the main file: e.g. neurosymbolic-modules/ec/bin/arc_simon.py (or another similarly structured file)
• To change the implementation of the primitives or implementations of tasks, edit neurosymbolic-modules/ec/dreamcoder/domains/arc/arcPrimitives.py and neurosymbolic-modules/ec/dreamcoder/domains/arc/makeTasks.py
• To change the implementation of the primitives in OCaml, edit neurosymbolic-modules/ec/solvers/program.ml. To recompile the OCaml: cd /om2/user/$USER/neurosymbolic-modules/ec; module add openmind/singularity; ./container.img make clean; ./container.img make;

If you get an a "dreamcoder.grammar.NoCandidates" exception, this means with the primitives you gave Dreamcoder, it can't find any programs which solve the task. Make sure the type defined for the tasks (for us, in dreamcoder/domains/arc/makeTasks.py) is what you expect, and that your primitives can be combined to give this type.

Example: if the type of the task is arrow(tinput, tgrid), but your functions are all arrow(tgrid, tgrid) type, you need to add a primitive with type arrow(tinput, tgrid).

If you get a JSON decode error when running compression, then there's probably a naming conflict with your ocaml/python primitives. For example, if you name a primitive "filter", but ocaml already had a primitive with that name used in one of the other domains. Try changing the names of your primitives to something unique, like "filter_list", or "arc_filter", and that should work.

If you get NaN's while running compression, and really weird primitives, check that your baseType("s") in python matches the make_ground "s" in ocaml. In the past we accidentally did baseType("tint") and make_ground "int".

We recently refactored the bidirectional code into the bidir-synth directory, so that our code is separate from the dreamcoder files. This directory also contains our RL code.

The recommended setup here is NOT singularity. We recommend an isolated Python 3.7+ environment (e.g. via virtualenv, conda, or poetry).

In your python environment of choice, install the needed python depenencies by running pip install -r bidir-synth/requirements.txt.

Tests have been consolidated into a single directory as well. To run the primitive and RL agent tests, go to the bidir-synth directory (cd bidir-synth) and run ./run_tests.sh.

Tests are autodiscovered from the bidir-synth/tests directory via Python unittest. Any file matching the pattern test*.py will be run as a test.

We also have a typechecking script for files in the bidir directory. We use python type annotations during program synthesis, so it is important that these type annotations are accurate. To run the typechecking script, go to the bidir-synth directory (cd bidir-synth) and run ./mypy.sh.

To run python scripts, go to the bidir-synth directory (cd bidir-synth) and run a command like

python -m rl.policy_gradient

The example above runs bidir-synth/rl/policy_gradient.py.

We are using MLflow to log experiments. To view experiments, run mlflow ui from the bidir-synth directory, then navigate to http://127.0.0.1:5000/#/, or wherever the command line says the server is listening. For more info, check out https://www.mlflow.org/docs/latest/quickstart.html. To check