Relational Neurosymbolic Markov Models (NeSy-MMs)

This repository contains the code for the paper "Relational Neurosymbolic Markov Models".

Installation

To install the required packages, run the following command:

pip install -r requirements.txt

Usage

Notice that the commands provided here are going to generate the dataset and then train the model, so it may take a while the first time. To generate the dataset in a different directory you can use the option --base_dir <path>. To use the Weights & Biases logging, you can use the option --wandb_mode "online". You can specify the seed with the option --seed <seed>. If you don't the default seed (42) is used.

Generative Experiment

To run the generative experiment, use the following commands, depending on what model you want to train. Training on an NVIDIA Tesla P100 GPU should take between 4 and 10 hours (per run, per model) depending on the model.

python nesy_mm/experiments/minihack_vae/run.py --grid_size 5
python nesy_mm/experiments/minihack_vae/run.py --grid_size 10 --downsample 2 --n_samples 20 --batch_size 5 --beta 50 --n_epochs 15

python nesy_mm/experiments/minihack_vae/baselines/deep_hmm/run.py --grid_size 5
python nesy_mm/experiments/minihack_vae/baselines/deep_hmm/run.py --grid_size 10 --downsample 2 --n_samples 20 --batch_size 5 --n_epochs 15

python nesy_mm/experiments/minihack_vae/baselines/transformer/run.py --grid_size 5
python nesy_mm/experiments/minihack_vae/baselines/transformer/run.py --grid_size 10 --downsample 2

To test the generative models, there is a Jupiter notebook generation.ipynb that you can use. This can be easily run on CPU.

Transition Learning Experiment

To run the transition learning experiment, use the following commands, depending on what model you want to train.

python nesy_mm/experiments/minihack_transition/run.py
python nesy_mm/experiments/minihack_transition/evaluate.py

python nesy_mm/experiments/minihack_transition/baselines/deep_hmm/run.py
python nesy_mm/experiments/minihack_transition/baselines/deep_hmm/evaluate.py

python nesy_mm/experiments/minihack_transition/baselines/transformer/run.py
python nesy_mm/experiments/minihack_transition/baselines/transformer/evaluate.py

In this case evaluation of the models is performed separately from the training script because we evaluate on multiple out-of-distribution datasets.