DeepProbLog

DeepProbLog provides several specialized modules for different stages of the machine learning and probabilistic reasoning pipeline:

• Data & Tensors: deepproblog.embeddings for vector representations, deepproblog.tensor for tensor operations, and deepproblog.dataset for managing training/evaluation data.
• Logic & Reasoning: deepproblog.query for formulating queries, deepproblog.solver for probabilistic inference, deepproblog.engine for the execution engine, and deepproblog.semiring for semiring-based reasoning.
• Neural Components: deepproblog.network for neural architectures and deepproblog.arithmetic_circuit for representing arithmetic operations within the logic framework.
• Training & Evaluation: deepproblog.optimizer for parameter updates, deepproblog.train for the training loop, and deepproblog.evaluate for assessing model performance.

The Forth example directory contains experiments inspired by the paper Programming with a Differentiable Forth Interpreter. These experiments demonstrate program induction using a differentiable Forth interpreter. There are three distinct experiments available:

1. Learning to add: Executes the addition learning task.
2. Learning to sort: Executes the sorting learning task.
3. Solving Word Algebra Problems (WAPs): Executes the WAP solving task.

Note: The data used in these experiments has been converted to the DeepProbLog format for compatibility, based on the original d4 repository.

# Learning to add
python Add/add.py

# Learning to sort
python Sort/sort.py

# Solving Word Algebra Problems (WAPs)
python WAP/wap.py

You can install the core DeepProbLog package using pip.

pip install deepproblog

Some MNIST experiments require pre-trained neural networks. To generate these models, execute the create_pretrained.py script located in the models/pretrained/ subdirectory.

python models/pretrained/create_pretrained.py

Install DeepProbLog using pip. Ensure your environment meets the core requirements: Python > 3.9, ProbLog, PySDD, PyTorch, and TorchVision.

pip install deepproblog

The Coins experiment requires Blender 2.79 to generate synthetic data. You can generate training and testing datasets by running the render.py script with the appropriate mode. If Blender is not in your system path, use the -b flag to specify the path to the Blender executable.

# Generate training data
python render.py train

# Generate testing data
python render.py test

# Generate data specifying a custom Blender path
python render.py train -b /path/to/blender

To verify that your installation is working correctly, install pytest and run the built-in test command via the deepproblog module.

pip install pytest
python -m deepproblog test

To generate the dataset for the Poker experiment, navigate to the data/ subdirectory and execute the render.py script.

Prerequisite: You must have inkscape installed on your system to run the rendering process.

cd data/
python render.py

To execute the MNIST addition experiments, run the addition.py script located in the example directory.

python addition.py

The WAP dataset used in the DeepProbLog examples is based on the work by Roy and Roth (2015). To set up the data for the Forth/WAP example, you must manually download the following files from the Roy and Roth fork repository:

• Full data: questions.json from data/commoncore_crossval/
• train.tsv: fold1.txt from data/commoncore_dev/
• dev.tsv: fold0.txt from data/commoncore_dev/
• test.tsv: fold0.txt from data/commoncore_test/

https://github.com/ninjin/roy_and_roth_2015/blob/master/data/commoncore_crossval/questions.json
https://github.com/ninjin/roy_and_roth_2015/blob/master/data/commoncore_dev/fold1.txt
https://github.com/ninjin/roy_and_roth_2015/blob/master/data/commoncore_dev/fold0.txt
https://github.com/ninjin/roy_and_roth_2015/blob/master/data/commoncore_test/fold0.txt

After downloading the raw JSON and text files, use the provided utility script data/wap_generate_data.py to process and generate the required .txt files for the experiment.

python data/wap_generate_data.py