Declarative DeepProbLog

A declarative extension for neuro-symbolic languages built on neural predicates. A neural predicate can be called with its image argument unbound; resolution then grounds it by sampling and decoding a prototype, so one trained model answers classification and generation queries alike.

Each top-level directory holds one host system: the vanilla baseline and our declarative version of it.

Directory	System	Vanilla	Declarative
deepproblog_examples/	DeepProbLog	mnist_class.py, mnistr_class.py, hwf_class.py	mnist_prototypes.py, mnist_n_prototypes.py, hwf_prototypes.py
deepstochlog_examples/	DeepStochLog	mathexpression.py, run_warcraft_pathfinding.py	mathexpression_prototype.py, warcraft_vae.py
neurasp_examples/	NeurASP	run_mnist_neurasp.py	run_mnist_declarative_neurasp.py
slash_examples/	SLASH	run_mnist_slash.py	run_mnist_declarative_slash.py
deepseaproblog_examples/	DeepSeaProbLog	run_mnist_dsp.py	—
vael_examples/	VAEL	run_mnist_vael.py	—
scallop_examples/	Scallop	run_mnist_scallop.py	—

Every runner accepts --help. See the README in each directory for its environment and commands.

Setup

DeepProbLog and DeepStochLog need deepproblog-dev (https://github.com/ML-KULeuven/deepproblog-dev), which ships the added predicates; follow its install guide, then:

pip install -r requirements.txt

The baselines each need their own environment; see their READMEs.